Agent Platform 2.0: Comprehensive Overview

This document is a consolidated reference covering the Agent Platform 2.0 end-to-end. It is designed to serve as input for generating training slides and presentation decks that provide both an executive-level overview and sufficient technical depth for a developer audience.

1. Elevator Pitch

Agent Platform 2.0 is an enterprise SaaS platform for building, deploying, and managing AI agents using ABL (Agent Blueprint Language) — the enterprise control plane for agentic AI, where deterministic governance meets autonomous reasoning. Instead of writing hundreds of lines of imperative framework code, teams describe what their agents should do in a structured, human-readable DSL — and the platform handles execution, multi-agent routing, 20+ communication channels, knowledge retrieval, guardrails, and full observability. Three sentences that capture the value:

Declarative, not imperative. A 15-line ABL definition replaces what would otherwise be several hundred lines of Python/JS framework boilerplate. Agent definitions are version-controlled, diffable, and readable by non-engineers.
Multi-agent orchestration is a first-class primitive. Supervisors route conversations to specialist agents with handoff, delegation, escalation, and fan-out patterns — all defined declaratively in ABL.
Enterprise-grade from day one. Multi-tenant isolation, encryption at rest/transit, input/output guardrails, audit logging, PII protection, and SOC 2 Type II compliance are built into the platform core — not bolted on.

2. Who Is This For?

Persona	Entry Point	What They Do
Agent Developer	Studio IDE / ABL DSL	Define agent behavior, write supervisor routing logic, configure tools and knowledge bases
QA / Tester	Evaluation Framework	Create personas, scenarios, and LLM judges to validate agent behavior at scale
Ops Engineer	Deployments & Sessions	Deploy agents across environments, monitor traces, track session analytics
Workspace Admin	Admin Console	Manage tenants, users, permissions, LLM credentials, channel connections, and security policies

3. Platform Architecture

Three Core Services

Studio — Browser-based IDE (Next.js, port 5173). No local installation. Design agents visually or in code, test interactively, deploy to production. Runtime — Stateless execution engine (Node.js, port 3112). Receives messages from 20+ channels, loads compiled agent definitions (IR), executes reasoning loops or step-based flows, calls LLMs, invokes tools, enforces guardrails, manages sessions, coordinates multi-agent handoffs. SearchAI — Full RAG pipeline (port 3113, with SearchAI Runtime on 3114). Handles document ingestion (PDF, DOCX, HTML, CSV, JSON), text extraction, chunking (semantic/fixed/hierarchical), embedding (BGE-M3), vector+keyword hybrid search, vocabulary resolution, knowledge graph extraction, and live data sync via connectors. Supporting Services — Admin Console (port 3003), Python services for document processing: Docling (8080), BGE-M3 embeddings (8000), Preprocessing (8003). Infrastructure: MongoDB 7.0+ (replica set), Redis 7+ (persistence, TLS), ClickHouse (optional, for analytics).

End-to-End Flow

4. ABL — The Agent Blueprint Language

Every enterprise deploying agents faces the same tension: you need AI autonomy to unlock value, but you need deterministic control to stay in production. ABL resolves this with a schema-driven language purpose-built for multi-agent orchestration.

Full-spectrum control — Not just autonomous or deterministic — ABL spans the entire control spectrum. Delegate autonomously, supervise selectively, or lock down as a deterministic state machine. One language, every mode.
Local + remote agent topology — Local agents run deterministic logic with creative reasoning. Remote agents connect any custom framework via A2A-enabled fan-out patterns. Your agents, your infrastructure, one orchestration layer.
Parallel execution with state semantics — Fan-out/fan-in orchestration with parent-child state transfer — so parallel agents stay coordinated without brittle glue code.
Compiled and immutable by design — Agent definitions compile into immutable artifacts. Every version is auditable, every deployment is reproducible, every change is governed.
Deep observability — native, not bolted on — Tracing across agent delegation chains is built into the runtime. No third-party instrumentation, no blind spots between handoffs.
AI-programmable platform — ABL is not just human-writable — it is designed as a code-generation target. AI authors agent blueprints; the platform compiles and enforces them.

File Types

Extension	Document Type	Purpose
`.agent.abl`	Agent definition	Most common — defines a single agent
`.supervisor.abl`	Supervisor	Top-level orchestrator for multi-agent
`.tools.abl`	Tool library	Reusable tool definitions shared across agents
`.agent.yaml`	YAML agent	Alternative YAML format for agent definitions
`BEHAVIOR_PROFILE`	Behavior profile	Reusable behavior configurations

Top-Level Sections

Every ABL agent can include these sections (only AGENT and GOAL are required):

Section	Purpose	Required
`AGENT`	Agent name (PascalCase_With_Underscores)	Yes
`GOAL`	Primary objective	Yes
`VERSION`	Semantic version (default: “1.0.0”)	No
`DESCRIPTION`	Human-readable summary	No
`LANGUAGE`	BCP 47 language code (e.g., “en”, “es-EC”)	No
`PERSONA`	Agent personality and communication style	No
`IDENTITY`	Structured identity (role, expertise, tone)	No
`LIMITATIONS`	Explicit boundaries (what agent cannot do)	No
`INSTRUCTIONS`	Operational procedural guidance	No
`EXECUTION`	Model & runtime configuration	No
`TOOLS`	Tool definitions and imports	No
`GATHER`	Information collection fields	No
`FLOW`	Structured execution steps	No
`MEMORY`	Session & persistent state	No
`CONSTRAINTS`	Business rule enforcement	No
`GUARDRAILS`	Input/output safety checks	No
`HANDOFF`	Agent transfer rules	No
`DELEGATE`	Sub-agent delegation	No
`ESCALATE`	Human escalation triggers	No
`COMPLETE`	Completion conditions	No
`NLU`	Natural language understanding config	No
`ON_START`	Session initialization	No
`ON_ERROR`	Error handlers	No
`HOOKS`	Lifecycle event handlers	No
`TEMPLATES`	Reusable response templates	No
`MESSAGES`	Customizable system-generated messages	No
`LOOKUP_TABLES`	Reference data for validation	No
`ATTACHMENTS`	File/media collection config	No
`SYSTEM_PROMPT`	Custom system prompt template	No

Anatomy of an ABL Agent

AGENT: Support_Assistant

VERSION: "1.0.0"

EXECUTION:
  model: claude-sonnet-4-5-20250929
  temperature: 0.3
  max_tokens: 4096
  max_reasoning_iterations: 15
  enable_thinking: true
  thinking_budget: 2048

GOAL: |
  Help customers with product questions. Be concise
  and friendly. If you do not know the answer, say so.

PERSONA: |
  Helpful product support assistant. Answers questions
  clearly and concisely.

LIMITATIONS:
  - "Cannot process payments or refunds"
  - "Cannot access customer account information"

TOOLS:
  search_knowledge(query: string) -> {results: object[], totalCount: number}
    description: "Search the product knowledge base"

GUARDRAILS:
  profanity_filter:
    kind: input
    check: not_contains_blocked_words(input)
    action: block
    message: "Please keep our conversation respectful."

INSTRUCTIONS: |
  1. Understand the customer's question
  2. Search the knowledge base for relevant information
  3. Provide a clear, sourced answer
  4. If unsure, offer to connect with a human agent

Execution Configuration

The EXECUTION block controls model selection and runtime behavior:

Property	Description
`model`	Primary LLM model identifier
`temperature`	Sampling temperature (0.0-1.0)
`max_tokens`	Maximum tokens in LLM response
`max_reasoning_iterations`	Maximum reasoning loop iterations
`max_flow_iterations`	Maximum step transitions in flow
`tool_timeout`	Tool execution timeout (ms)
`llm_timeout`	LLM inference timeout (ms)
`session_idle_timeout`	Session expiration timeout
`fallback_model`	Model for unavailability fallback
`enable_thinking`	Extended thinking capability
`thinking_budget`	Token budget for extended thinking
`compaction_threshold`	Context usage ratio triggering auto-compaction
`inline_gather`	Collect gather fields inline during conversation
`voice_latency_target`	Target latency for voice responses

Per-Operation Model Routing — Route different operations to different models within the same agent:

EXECUTION:
  models:
    extraction: gpt-4o-mini        # Fast model for data extraction
    response_gen: claude-sonnet     # Balanced model for responses
    reasoning: claude-opus          # Powerful model for complex reasoning
    coordination: gpt-4o-mini      # Fast model for multi-agent routing

One Agent, Two Execution Modes

Every agent reasons by default. When you need structured execution, add a FLOW section with steps. This is not a different type of agent — it is the same agent with an optional flow definition.

Feature	Agent (default)	Agent with steps
Decision making	LLM decides next action autonomously	Steps define explicit transitions
Best for	Open-ended conversations, complex reasoning	Structured workflows, deterministic processes
Reasoning control	Always on	Per-step (`REASONING: true/false`)
Editor UI	Full-page configuration panel	Full-page config + visual flow canvas

Per-Step Reasoning Control — The Differentiator

Within a FLOW, each step’s REASONING toggle controls whether it uses LLM reasoning or runs deterministically:

FLOW:
  collect_preferences:
    REASONING: false         # Deterministic data collection
    GATHER:
      - destination: required
      - travel_dates: required
    THEN: research_options

  research_options:
    REASONING: true          # LLM autonomously researches
    GOAL: "Research travel options for {{destination}}"
    AVAILABLE_TOOLS: [search_flights, search_hotels, get_weather]
    MAX_TURNS: 8
    EXIT_WHEN: "all options found"
    STEP_CONSTRAINTS:
      - REQUIRE total_cost <= budget
    THEN: present_plan

  present_plan:
    REASONING: false         # Deterministic presentation
    RESPOND: "Here is your travel plan: {{compiled_itinerary}}"
    THEN: COMPLETE

This gives teams deterministic control where they need predictability (data collection, confirmations, API calls) and LLM autonomy where they need flexibility (research, analysis, open-ended Q&A).

Flow Control Constructs

Construct	Purpose
`THEN`	Basic next-step transition
`ON_RESULT`	Multi-way branching based on tool result (IF/ELSE)
`ON_INPUT`	Branching based on user input after GATHER
`ON_SUCCESS/FAIL`	Alternative to ON_RESULT for binary branching
`DIGRESSIONS`	Intent-based flow interruptions with resume/goto
`SUB_INTENTS`	Step-scoped intent patterns for corrections
`MAX_ATTEMPTS`	Limit step executions with ON_EXHAUSTED fallback
`CALL...WITH...AS`	Invoke tool with parameters, store result
`SET`	Assign session variables
`CHECK`	Evaluate condition with ON_FAIL
`CLEAR`	Remove variables from context
`TRANSFORM`	Filter, map, sort, limit arrays
`PRESENT`	Display formatted presentation before collection
`COMPLETE`	Exit the flow

Interactive Actions in Flow

Steps can present interactive UI elements (buttons, select menus, input fields) and handle user responses:

  confirm_booking:
    REASONING: false
    RESPOND: "Review your booking details:"
    ACTIONS:
      - id: confirm
        type: button
        label: "Confirm Booking"
        value: confirmed
      - id: cancel
        type: button
        label: "Cancel"
        value: cancelled
    ON_ACTION:
      - ACTION_ID: confirm
        SET:
          booking_status: "confirmed"
        THEN: process_payment
      - ACTION_ID: cancel
        RESPOND: "Booking cancelled."
        THEN: COMPLETE

Expressions & Built-in Functions

ABL includes a full expression language with 36+ built-in functions: Operators: ==, !=, >, <, >=, <=, IN, NOT IN, contains, matches, AND, OR, NOT, IS SET, IS NOT SET, EXISTS, EMPTY Math: ADD, SUB, MUL, DIV, ROUND, ABS, MIN, MAX String: UPPER, LOWER, TRIM, SUBSTRING, REPLACE, SPLIT, JOIN, PAD_START, PAD_END, REPEAT Formatting: MASK (last4, first4, N*N patterns), FORMAT_CURRENCY, FORMAT_DATE, ORDINAL Type: IS_ARRAY, IS_NUMBER, IS_STRING, TO_NUMBER, TO_STRING, LENGTH Array/Object: ARRAY_FIND, ARRAY_FIND_INDEX, OBJECT_KEYS, OBJECT_VALUES, OBJECT_MERGE Utility: COALESCE, NOW, UNIQUE_ID Template Syntax: {{variable_name}}, {{#if variable}}...{{/if}}, {{FORMAT_CURRENCY(balance, "USD")}}

Data Types

Primitive types: string, number, boolean, date, datetime Complex types: array<T>, object<{...}>, enum<[...]>, union<[...]>, nullable<T> Lookup Tables — Reference data for validation with three source types:

Source	Description
`inline`	Static values defined directly in ABL
`collection`	External data collection (database-backed)
`api`	HTTP endpoint returning lookup data

Lookup tables support case_sensitive, fuzzy_match, and fuzzy_threshold (default 0.85) for flexible matching.

Attachments

GATHER can collect files and media with automatic processing:

ATTACHMENTS:
  receipt_photo:
    prompt: "Please upload a photo of your receipt"
    category: image
    required: true
    max_file_size_mb: 10
    allowed_mime_types: ["image/jpeg", "image/png"]
    ocr_enabled: true

Category	Processing	Access Pattern
`document`	OCR text extraction	`{{field.extracted_text}}`
`image`	OCR text extraction	`{{field.extracted_text}}`
`audio`	Speech-to-text transcription	`{{field.transcript}}`
`video`	Keyframe extraction + optional transcription	`{{field.keyframes}}`

5. Multi-Agent Orchestration

ABL has first-class support for multi-agent systems with four coordination patterns:

Orchestration Patterns

Pattern	Description	User Experience
Handoff	Route conversation to a specialist agent	User is “transferred” to a new agent
Delegation	Send a sub-task to an agent and get results back	Transparent — user doesn’t see the delegation
Escalation	Transfer to a human agent with full context	User gets connected to a real person
Fan-out	Run multiple agents in parallel, merge results	Transparent — user sees a single combined result

Session Hierarchy

Context flow between agents:

Data Type	Transferred?
Session metadata	Yes (non-internal values forwarded)
Conversation history	Configurable: none/summary_only/full/last_N
GATHER progress	No (not transferred)
Custom variables (SET)	Yes (transferred as metadata)
Agent-specific state	No (not transferred)
Persistent memory	Via `grant_memory` paths

Supervisor Definition

SUPERVISOR: Travel_Supervisor

GOAL: "Route customers to the right specialist"

AGENTS:
  - REF: Flight_Search
    ALIAS: flights
    CAPABILITIES: ["flight booking", "flight status"]
  - REF: Hotel_Search
    ALIAS: hotels
    CAPABILITIES: ["hotel booking", "accommodation"]

HANDOFF:
  - TO: Flight_Search
    WHEN: intent contains "flight" OR intent contains "fly"
    PRIORITY: 1
    CONTEXT:
      pass: [destination, date, passengers]
      summary: "User needs flight booking assistance"
      history: summary_only
      grant_memory: ["user.preferences.travel"]
    RETURN: true
    ON_RETURN: "Summarize the booking result"
    MAP:
      booking_id: flight_booking_id

  - TO: Hotel_Search
    WHEN: intent contains "hotel" OR intent contains "stay"
    PRIORITY: 2
    CONTEXT:
      pass: [destination, checkin, checkout, guests]
    RETURN: true

  - TO: Live_Agent_Transfer
    WHEN: user.frustration_detected == true
    PRIORITY: 0
    CONTEXT:
      pass: [conversation_summary, transfer_reason]
    RETURN: false

Supervisor Advanced Features

ROUTING — Priority-ordered conditional routing rules with flags and constraints:

ROUTING:
  - NAME: "High-value routing"
    PRIORITY: 1
    WHEN: customer_tier == "platinum"
    THEN: ROUTE_TO premium_agent
    FLAGS: [priority_boost]
    CONSTRAINTS:
      channels: ["voice", "chat"]

STATE — Typed state schema for supervisor-managed variables (organized by namespace, with source tracking) POLICIES — High-level behavioral rules with allowedWhen, forbiddenWhen, and triggerSignal conditions BEHAVIOR — Supervisor capabilities: canRespondDirectly, allowedDirectActions, forbiddenActions COMMUNICATION — Language, formality level (formal/informal/neutral), pronouns, vocabulary, constraints

Delegation (Call-and-Return)

DELEGATE:
  - AGENT: Price_Calculator
    WHEN: "user asks about pricing"
    PURPOSE: "Calculate total price with discounts"
    INPUT:
      items: cart_items
      membership: customer_tier
    RETURNS:
      total_price: calculated_total
      discount_applied: discount_percent
    USE_RESULT: "Present the pricing to the user"
    TIMEOUT: 30000
    ON_FAILURE: respond
    FAILURE_MESSAGE: "Unable to calculate price right now."

Fan-out (Parallel Delegation)

Multiple DELEGATE entries with overlapping WHEN conditions execute in parallel. Each has independent TIMEOUT and ON_FAILURE handling, with distinct variable names in RETURNS.

Escalation to Humans

ESCALATE:
  triggers:
    - WHEN: "user.frustration_detected == true"
      REASON: "Customer expressing frustration"
      PRIORITY: high
      TAGS: ["customer-experience"]
    - WHEN: "handoff_count >= 3"
      REASON: "Multiple failed handoffs"
      PRIORITY: critical
  context_for_human: [conversation_summary, customer_tier, account_id]
  routing:
    queue: "tier2-support"
    skill_tags: ["billing", "retention"]
  on_human_complete:
    - IF: "outcome == 'resolved'"
      THEN: COMPLETE
    - IF: "outcome == 'needs_agent'"
      THEN: HANDOFF Specialist_Agent

Human Task System

When escalation triggers, a human task appears in the team inbox:

Property	Description
Task types	approval, data_entry, review, decision, escalation
Status	pending, assigned, in_progress, completed, expired, cancelled
Priority	low, medium, high, critical
Form fields	text, number, boolean, select, textarea, date
SLA	Deadline with escalation chains

Sessions suspend until human resolution, then resume with human.field context available.

Project Orchestration Patterns (Pre-built Topologies)

Pattern	Description	Best For
Concierge	Single front-facing agent delegates behind the scenes	Customer service, help desk
Router	Triage agent routes to specialists directly	Multi-department support
Tiered	Triage -> L1 -> L2 with escalation paths	Technical support
Custom	Flexible topology	Advanced use cases

Real-World Scale: Jupiter Banking Example

A 12-agent enterprise banking deployment with priority-ordered routing:

6. Studio — The Development Environment

Studio is a browser-based IDE — no local installation required. It covers the entire agent lifecycle.

Key Capabilities

Project Dashboard — Grid of project cards with metrics (agents, sessions, tokens, cost, containment rate), search, and “New Project” dropdown
Agent Editor — 17 configuration sections organized into six groups:
- Identity & Core: Identity, Execution, Tools
- Data & Logic: Gather, Memory, Flow
- Safety & Behavior: Constraints, Guardrails, Behavior
- Coordination: Handoffs, Delegates, Escalation
- Lifecycle: On Start, Error Handling, Completion
- Advanced: Templates, Definition
Visual Flow Editor — Canvas-based flow designer with drag-and-drop step palette, node connections, zoom/pan, and property panel for editing step details
Code Editor — Monaco-powered ABL editor with syntax highlighting and real-time validation; changes sync between visual and code modes
AI Architect — Context-aware AI assistant built into Studio (see details below)
AI-Guided Project Creation — Multi-phase wizard (see details below)
Test Chat — Interactive testing with real-time trace inspection, debug panel, session management, and variable state viewer
Command Palette — Cmd+K (Mac) / Ctrl+K quick navigation across pages, agents, and actions
Project Switcher — Dropdown at top of sidebar to switch projects without returning to dashboard
Agent Versioning — Version history with status tracking (draft, testing, staged, active, deprecated), side-by-side diff viewer, and promotion workflow
Pattern-Aware Agent Creation — Auto-assigns roles (Concierge, Router, Tiered, Custom) based on selected orchestration pattern
WebSocket Connectivity Indicator — Real-time connection status in header

AI Architect

A context-aware AI assistant embedded in Studio that helps design, build, and improve agents:

Context Awareness: Adapts to current page (Agents, Sessions, Overview), current agent, and current section being edited
Suggestion Chips: Context-specific suggestions:
- Agent editor: “Explain code”, “Add error handling”, “Suggest improvements”, “Generate tests”
- Session viewer: “Analyze session”, “Suggest fix”
- Project overview: “Summarize health”, “Identify bottlenecks”
Code Changes: Proposes changes via diff view with Accept/Reject/Refine workflow
Conversation History: Project-scoped, persisted to server across sessions
Controls: Maximize/restore, minimize, close; accessible from header icon, section buttons, or floating button

AI-Guided Project Creation Wizard

A multi-phase process for generating complete project structures:

Welcome — Introduction to the wizard
Interview — AI asks targeted questions about your use case
Upload — Analyze uploaded documents (product manuals, FAQs, policies)
Generating — AI generates project structure, agents, tools, and routing
Reveal — Preview generated structure
Review — Edit and refine before creation
Create — Generate the project with all components

Section	Pages
Build	Overview, Agents, Workflows
Resources	Tools, Knowledge Bases, Integrations
Evaluate	Evaluations, Experiments
Operate	Sessions, Deployments, Inbox, Alerts, Transfer Sessions
Insights	Dashboard, Agent Performance, Quality Monitor, Customer Insights, Voice Analytics
Govern	Guardrails, Governance
Settings	Members, API Keys, Models, Config Variables, Git, Runtime Config, Trace Dimensions, Agent Transfer, PII Protection

7. Tools & Integrations

Agents connect to external systems through five tool types:

Tool Types

Type	Description	Use Case
HTTP Tools	REST API calls with auth, retry, circuit breaker	CRM lookup, payment processing, booking APIs
Code Sandbox	Isolated JavaScript/Python execution	Custom calculations, data transforms, ML inference
MCP Servers	Model Context Protocol connections	Standardized multi-tool servers (browser, database, etc.)
Lambda Tools	Serverless function invocations	Lightweight compute, event processing
Async Webhooks	Suspend/resume for long-running operations	Payment processing, approval workflows, async tasks

Tool Declaration Syntax

TOOLS:
  tool_name(param1: type, param2: type = default) -> {field: type, field?: type}
    description: "What this tool does"
    type: http | mcp | sandbox | lambda | async_webhook
    # ... binding-specific properties

Tool Execution Hints

  hints:
    cacheable: true          # Results can be cached
    latency: fast|medium|slow
    side_effects: true       # Modifies external state
    requires_auth: true
    timeout: 5000

Key Tool Features

Authentication: None, API key, Bearer, OAuth2 client credentials, OAuth2 user authorization, SAML, custom headers
OAuth2 Details: token_url, client_id, client_secret, scopes, provider (for consent UI); automatic token caching and refresh
Reliability: Configurable timeout, retry count, retry_delay, exponential backoff, circuit breaker (threshold + reset_ms)
Rate Limiting: Per-tool requests/second caps
Confirmation Prompts: Require user approval before executing tools with side effects (require: always|never|when_side_effects, immutable_params for protected values)
Async Webhooks: Platform suspends session, registers callback URL, resumes on webhook callback with HMAC-SHA256 signature verification; configurable timeout (supports minutes to hours)
Result Mapping: on_result for variable assignment from tool output; on_error for error handling with variable assignment
PII Access Control: pii_access levels (tools, user, logs, llm) control where sensitive data flows
Context Access: context_access with read and write permissions for session variables
Reusable Tool Files: .tools.abl files with base_url, shared auth, timeout, retry, headers defaults; import syntax: FROM "./tools/file.tools.abl" USE: tool1, tool2

Code Sandbox Tools

TOOLS:
  calculate_tax(amount: number, state: string) -> {tax: number, total: number}
    description: "Calculate sales tax"
    type: sandbox
    runtime: javascript    # or python
    timeout: 5000
    memory_mb: 128
    code: |
      function calculate_tax({ amount, state }) {
        const rates = { CA: 0.0725, NY: 0.08, TX: 0.0625 };
        const tax = amount * (rates[state] || 0.05);
        return { tax: Math.round(tax * 100) / 100, total: amount + tax };
      }

Tool Definition Example (HTTP)

TOOLS:
  process_payment(amount: number, currency: string, card_token: string) -> {transaction_id: string, success: boolean}
    description: "Process a payment"
    type: http
    endpoint: "https://payments.example.com/v1/charge"
    method: POST
    auth: bearer
    timeout: 10000
    retry: 3
    retry_delay: 1000
    circuit_breaker:
      threshold: 5
      reset_ms: 60000
    confirmation:
      require: always
      immutable_params: [amount, currency]

Error Handling

Agent-level and step-level error handlers with retry and backoff:

ON_ERROR:
  tool_timeout:
    RESPOND: "The service is taking longer than expected."
    RETRY: 2
    RETRY_DELAY: 3000
    RETRY_BACKOFF: exponential
    RETRY_MAX_DELAY: 30000
    THEN: CONTINUE
  tool_error:
    RESPOND: "Something went wrong. Let me try a different approach."
    THEN: ESCALATE

Error types: tool_timeout, tool_error, validation_error, llm_error, routing_failure, agent_unavailable, timeout. Each supports subtypes for fine-grained matching and BACKTRACK_TO for flow step recovery.

8. Knowledge Bases & RAG

SearchAI provides an integrated RAG pipeline with no custom retrieval code required.

Ingestion Pipeline

Upload/Connect -> Ingest -> Extract Text -> Chunk -> Enrich -> Embed -> Store

Ingest — Register document, check for duplicates via content hashing
Extract — Format-specific text extraction (PDF, DOCX, JSON, HTML, plain text)
Chunk — Split into searchable segments
Enrich — Entity detection, summary generation, language detection, knowledge graph extraction
Embed — Vector generation via BGE-M3 (multilingual)
Store — Vector database with metadata and permissions

Supported Formats

Documents: PDF, DOCX, TXT, Markdown
Web: HTML pages, crawls
Structured: JSON, CSV
Rich media: Images in documents (with OCR)

Chunking Strategies

Strategy	Description
Fixed-size	Target token size with configurable overlap
Semantic	Splits on natural boundaries (paragraphs, sections)
Hierarchical	Tree of summarized chunks
Sliding window	Overlapping windows for context preservation

Search Strategies

Strategy	Description
Hybrid (recommended)	Combines vector similarity + keyword matching
Semantic	Pure vector search for conceptual matching
Structured	Metadata-filtered queries for structured data
Aggregation	Sum, avg, count, min, max over structured datasets

Agent-Level Search Tools

Tool	Purpose
`search_hybrid`	Vector + keyword search (default)
`search_vector`	Pure semantic search
`vocabulary_resolve`	Business terms to metadata filters
`search_aggregate`	Sum, avg, count, min, max over data
`search_structured`	Metadata filter queries
`search_list`	Paginated structured results

Search Configuration

topK: Number of results (default 5)
minScore: Confidence threshold (default 0.7)
strategy: hybrid/vector/structured/aggregation
cacheTtl: Embedding cache TTL

Live Data Connectors

Connector	What It Syncs	Auth Type
SharePoint	Documents and pages	OAuth 2.0
Confluence	Pages and blog posts	API token
Jira	Issues, comments, attachments	API token
Salesforce	Knowledge articles, cases	OAuth 2.0
ServiceNow	Knowledge articles, incidents	OAuth 2.0 / API key
HubSpot	KB articles	API key
Web Crawler	Any website	Configurable crawl rules
Database	SQL query results	Connection string

All connectors support delta sync on configurable schedules — initial full sync followed by periodic delta syncs. Deleted document removal is automatic.

9. Data Collection with GATHER

GATHER collects structured data from users through natural conversation.

Key Features

LLM Extraction: Users provide multiple fields in a single message; the LLM extracts them all
Strategies: llm (default — LLM-based), pattern (regex-based), hybrid (combined)
Field Types: string, number, date, email, phone, boolean
Validation: Regex patterns (REGEX), code-based (CODE), LLM-based judgment (LLM), cross-field validation via CONSTRAINTS
Inference: Auto-extract values from conversation context with configurable infer_confidence threshold (default 0.8) and optional infer_confirm
Corrections: Users can change previously provided values without restarting (enable with CORRECTIONS: true)
Progressive Activation: Fields activate only after dependencies are met (activation: progressive, depends_on)
Extraction Hints: Guide LLM on how to interpret ambiguous input
List Collection: list: true to collect arrays of values
Range Collection: range: true to collect {low, high} pairs
Preferences: preferences: true to categorize into accept/desire/avoid/refuse
Sensitive Fields: sensitive: true with sensitive_display (redact/mask/replace) and mask_config
Transient Fields: transient: true to auto-clean after GATHER completes
Custom Retry: Per-field retry_prompt and max_retries

GATHER Example

GATHER:
  FIELDS:
    destination:
      prompt: "Where would you like to travel?"
      type: string
      required: true
      infer: true
      infer_confidence: 0.9
      extraction_hints: "Look for city names, country names, or airport codes"

    budget:
      prompt: "What's your budget range?"
      type: number
      range: true
      validate: min(100)
      depends_on: [destination]
      activation: progressive

    travel_dates:
      prompt: "When do you want to travel?"
      type: date
      required: true
      validate: "value > NOW()"
      retry_prompt: "Please provide a future date."

  STRATEGY: llm
  CORRECTIONS: true
  COMPLETE_WHEN: "destination IS SET AND travel_dates IS SET"

10. Memory & State Management

Session Variables

Scoped to a single conversation; created at session start, discarded at end
Declared in MEMORY.session section
Types: string, number, object, array, boolean
Properties: TYPE, DESCRIPTION, INITIAL, RESET (per_session / per_step / never)
Used in templates ({{var}}), conditions, and tool calls

State Manipulation Directives

# SET -- assign values (literals, expressions, function calls)
SET:
  total: ADD(subtotal, tax)
  greeting: "Hello, {{customer_name}}"
  order_id: UNIQUE_ID(12)
  timestamp: NOW()

# CLEAR -- remove variables from session
CLEAR: [temp_result, intermediate_data]

# TRANSFORM -- filter, map, sort, limit arrays
TRANSFORM:
  SOURCE: search_results
  AS: item
  INTO: filtered_results
  FILTER: "item.score >= 0.8"
  MAP: "{title: item.title, url: item.url}"
  SORT_BY: "item.score DESC"
  LIMIT: 5

Persistent Memory

Survives across sessions (user preferences, history)
Dot-notation paths: user.preferences.language, project.faq_count
Scope: user (per-user) or project (shared across users)
Access control: read, write, readwrite
TTL-based expiration (s/m/h/d)
Stored in fact store, accessible via Memory API

Auto-Recall (Remember/Recall Pattern)

MEMORY:
  remember:
    - WHEN: "preferred_language IS SET"
      STORE:
        preferred_language -> user.preferences.language
      TTL: 90d

  recall:
    - ON: session:start
      ACTION: inject_context
      PATHS: ["user.preferences"]
    - ON: search:before
      ACTION: load_memory
      DOMAIN: "user.history"
    - ON: session:start
      ACTION: prompt_llm
      INSTRUCTION: "Recall the user's previous interactions and preferences"

11. Safety & Guardrails

Multi-layered safety built into the platform core with a three-tier architecture:

Three-Tier Guardrail Implementation

Tier	Type	Speed	Description
1	CEL-based	Fastest	Deterministic pattern matching (regex)
2	Model-based	Fast	Pre-trained safety models (moderation)
3	LLM-based	Moderate	Natural language LLM judgment checks

Input Guardrails

Pattern-based detection (regex for PII: SSN, credit cards, phone numbers)
Model-based content moderation (OpenAI Moderation, Azure Content Safety, Lakera Guard, Anthropic)
LLM-based topic relevance checks
Actions: block, warn, redact, escalate, reask, filter
Priority ordering (lower numbers run first)
severity_actions for per-severity branching

Output Guardrails

PII leak prevention
Toxicity scoring
Response length limits
Fix strategies: strip_html, redact_pii, truncate, normalize, custom (CEL expression)
max_reasks for reask action (default 2)
filter_min_length for minimum content length after filtering

Streaming Guardrails

streaming: true to enable real-time evaluation during response generation
streaming_interval: token, sentence, or chunk_size
Stops generation mid-stream on violation

Business Constraints

CONSTRAINTS:
  booking_rules:
    - REQUIRE passenger_count <= 9
      ON_FAIL: "We can only book for up to 9 passengers."

    - REQUIRE departure_date > NOW()
      ON_FAIL:
        RESPOND: "Travel dates must be in the future."
        COLLECT: [departure_date]
        THEN: retry

  risk_controls:
    - LIMIT daily_transactions <= 50
      ON_FAIL: ESCALATE

    - RESTRICT destination IN ["sanctioned_countries"]
      ON_FAIL: BLOCK

Constraint keywords: REQUIRE (assertion), LIMIT (numeric boundary), RESTRICT (prohibition) Constraint labels: always, pre_search, pre_action, pre_booking, or custom labels. They group related rules for readability; use WHEN or structural BEFORE for runtime gating.

Guardrail Providers

Provider	Type	Categories
OpenAI Moderation	Model-based	hate, violence, sexual, self-harm, harassment
Azure Content Safety	Model-based	hate, violence, sexual, self-harm
Anthropic	Model-based	harmless, honest
Lakera Guard	Model-based	prompt injection, PII, toxic
Built-in PII	Pattern-based	SSN, credit cards, phone, email
Custom HTTP/Webhook	Custom	Any via external endpoint
Custom LLM	LLM-based	Any via natural language check

Guardrail Policies (Centralized)

Project-level and agent-level scope (project policies are universal, agent policies are domain-specific)
Settings: enableInputGuardrails, enableOutputGuardrails, enableStreamingGuardrails
Budget controls: maxEvalsPerMinute, maxCostPerDay
Caching: enabled, ttlSeconds to reduce duplicate evaluations
Constitution principles for organizational safety guidelines
Audit trail of all guardrail evaluations
Fail modes: fail-open (flag if provider unavailable) or fail-closed (block)

Built-in Guardrail Templates

Pre-configured templates for common patterns: account_number_masking, credential_input, ssn_protection, profanity_filter, harmful_content_detection

Rate Limiting

Level	Controls
Tenant	requests/min, messages/min
Project	messagesPerMinute, sessionsPerHour, tokensPerMinute
Per-user	Individual limits
Token	tokensPerMinute, tokensPerDay budgets
SDK channel	messagesPerMinute, maxConcurrentSessions

Rate limit responses: HTTP 429 with retryAfter header.

Audit Logging

Enabled by default; tracks mutations (create, update, delete) and auth events
Fields: timestamp, actor, action, resource, outcome, requestId, ipAddress
Request/response body capture with redactFields for sensitive data
Export to CSV or external SIEM systems
Retention: 90+ days on Professional/Enterprise plans

12. NLU & Multi-Language Support

Intent Classification

NLU:
  intents:
    - NAME: book_flight
      PATTERNS: ["book", "flight", "fly to"]
      EXAMPLES:
        - "I need to book a flight to London"
        - "Can you find me flights next Tuesday?"
      ENTITIES: [destination, travel_date]

    - NAME: check_status
      PATTERNS: ["status", "where is", "track"]
      EXAMPLES_FILE: "./intents/status-examples.txt"

  entities:
    - NAME: destination
      TYPE: enum
      VALUES: ["London", "Paris", "Tokyo"]
      SYNONYMS:
        London: ["LON", "Heathrow", "LHR"]

    - NAME: travel_date
      TYPE: date
      VALIDATION: "value >= NOW()"

Pattern-based and example-based intent definitions
Entity types: enum, pattern (regex), date, number, location, free_text
Categories for high-level routing
Synonyms for canonical value mapping

Embedding-Based Semantic Matching

NLU:
  embeddings:
    enabled: true
    provider: openai
    model: text-embedding-3-small
    threshold: 0.75
    cache_ttl: 3600

Provides robust classification when patterns alone are insufficient.

NLU Model Configuration

fast: Low-latency model for routing (e.g., gpt-4o-mini)
balanced: Higher-accuracy model for extraction (e.g., claude-sonnet)
Evaluation: log_predictions, ab_test, confidence_threshold
External NLU: config_file reference for external YAML configuration

Multi-Language Support

ISO 639-1 language codes with auto-detection
allow_code_switching: Follow user’s language changes mid-conversation
Language-specific model routing via language_models
Domain glossary for consistent terminology across languages
Multilingual embeddings (BGE-M3) for cross-language search
Language-based agent routing via handoff conditions
TEMPLATES for multi-language pre-written responses
Environment variable support for deployment-specific language overrides

13. Lifecycle & Hooks

ON_START (Session Initialization)

Fires before any user input; used for welcome messages, variable setup, and tool calls:

ON_START:
  SET:
    session_start: NOW()
    greeting_sent: false
  CALL: get_user_profile
    WITH: {user_id: "{{session.userId}}"}
    AS: user_profile
  RESPOND: "Welcome back, {{user_profile.name}}! How can I help?"
  VOICE:
    INSTRUCTIONS: "Speak warmly and professionally"
  RICH_CONTENT:
    ADAPTIVE_CARD: |
      { "type": "AdaptiveCard", ... }
  ACTIONS:
    - id: quick_start
      type: button
      label: "Quick Start Guide"

HOOKS (Lifecycle Event Handlers)

HOOKS:
  before_agent:
    SET: {agent_activated: NOW()}
    CALL: log_activation
  after_agent:
    CALL: save_summary
  before_turn:
    CALL: check_session_validity
  after_turn:
    SET: {last_activity: NOW()}

Hook	Fires When
`before_agent`	Agent first activated
`after_agent`	Agent completes or deactivated
`before_turn`	Before processing user message
`after_turn`	After turn completes

TEMPLATES (Reusable Responses)

TEMPLATES:
  booking_confirmation:
    DEFAULT: "Your booking #{{booking_id}} is confirmed for {{date}}."
    MARKDOWN: "## Booking Confirmed\n**ID**: {{booking_id}}\n**Date**: {{date}}"
    ADAPTIVE_CARD: |
      { "type": "AdaptiveCard", "body": [...] }
    VOICE:
      INSTRUCTIONS: "Read the confirmation slowly and clearly"
    ACTIONS:
      - id: view_details
        type: button
        label: "View Details"

Reference with TEMPLATE(booking_confirmation) in RESPOND or ON_START.

MESSAGES (System Message Customization)

Override system-generated messages for consistent branding:

error_default — Generic error message
constraint_blocked — Constraint violation message
escalation_format — Human escalation message
conversation_complete — Session end message
gather_prompt — Data collection prompt

14. Rich Content & Voice

Rich Content Formats

Format	Channel	Description
`MARKDOWN`	Web, general	Formatted markdown
`ADAPTIVE_CARD`	Microsoft Teams	Adaptive Cards JSON
`SLACK`	Slack	Block Kit JSON
`WHATSAPP`	WhatsApp	Interactive messages (buttons, lists)
`AG_UI`	AG-UI / CopilotKit	AG-UI events
`HTML`	Email, web	HTML content
`CAROUSEL`	Web, mobile	Scrollable card collection

Carousels

RICH_CONTENT:
  CAROUSEL:
    - title: "Flight Option 1"
      subtitle: "$299 - Direct"
      imageUrl: "https://example.com/flight1.jpg"
      buttons:
        - id: select_flight_1
          type: button
          label: "Select"
          value: "FL001"
    - title: "Flight Option 2"
      subtitle: "$199 - 1 Stop"
      imageUrl: "https://example.com/flight2.jpg"
      buttons:
        - id: select_flight_2
          type: button
          label: "Select"
          value: "FL002"

Interactive Actions

ACTIONS:
  - id: departure_select
    type: select
    label: "Select departure city"
    options: ["New York", "Los Angeles", "Chicago"]
  - id: passenger_count
    type: input
    inputType: number
    label: "Number of passengers"
    placeholder: "1-9"
    required: true
  - id: submit_search
    type: button
    label: "Search Flights"
    value: search

submitLabel: "Submit"
submitId: search_form

Voice (SSML)

VOICE:
  ssml: |
    <speak>
      Your flight <say-as interpret-as="characters">UA</say-as>
      <say-as interpret-as="cardinal">472</say-as>
      departs at <say-as interpret-as="time">3:45pm</say-as>.
      <break time="500ms"/>
      Gate <say-as interpret-as="characters">B</say-as>
      <say-as interpret-as="cardinal">12</say-as>.
    </speak>
  instructions: "Speak clearly and at a measured pace"
  plain_text: "Your flight UA 472 departs at 3:45 PM from Gate B12."

File Attachments

ATTACHMENTS block declares accepted file types per category (image, document, audio, video)
Automatic processing: OCR for images/documents, transcription for audio, keyframe extraction for video
Max file size, allowed MIME types configurable
Channel handling normalized across web, Slack, WhatsApp, Teams, Email, Telegram
Upload API: POST /api/projects/:projectId/sessions/:id/attachments
Status tracking: scanStatus, processingStatus, embeddingStatus

15. Testing & Evaluation Framework

A complete testing pipeline built into the platform:

Components

Component	Purpose
Personas	Synthetic user profiles with communication style, domain knowledge, behavioral traits, goals
Scenarios	Test conversation scripts with initial messages, expected outcomes, milestones
Evaluators	LLM judges, trajectory analyzers, code scorers, human review
Eval Sets	Bundles of personas + scenarios + evaluators as reusable test suites
Runs	Execute the Cartesian product: every persona x scenario x variant, scored by all evaluators

Evaluation Pipeline

Test Data Generation -> Conversation Execution -> LLM Judge Evaluation -> Recommendation Generation

Studio Test Chat

Interactive testing with real-time message input/response
Observe tool calls, flow transitions, handoffs, guardrails in real-time
Metadata per message: step executed, tools called, response time
Trace inspection: span tree, decision points, timing
Session management: start new, view history, copy session ID
Variable state viewer: inspect current session variables

Test Personas

Communication styles: casual, formal, technical, terse, verbose
Domain knowledge levels: beginner, intermediate, expert
Behavioral traits and goals
AI-generated personas for quick setup
Adversarial personas: prompt_injection, social_engineering, off_topic, abusive, edge_case

Test Scenarios

Category, difficulty (easy/medium/hard), entry agent
Initial message and expected outcome
Max turns for conversation length
Expected milestones: key checkpoints the conversation should hit
Agent path: expected sequence of agents (for multi-agent testing)
Tags for filtering and organization

Evaluator Types

Type	Description
LLM Judge	LLM scores conversations using rubric (1-5 or pass/fail)
Trajectory Evaluator	Assesses execution path: milestones, handoffs, efficiency
Code Scorer	Deterministic checks via regex, keywords, thresholds
Human Review	Manual assessment by human reviewer

LLM Judge Features

Categories: quality, safety, efficiency, empathy, tool_correctness, custom
Configurable judge model (gpt-4o, claude-sonnet, etc.) and temperature
Chain of thought: explain reasoning before scoring
Bias mitigation: Position swap, blind evaluation, cross-model judge, evidence-first mode

Eval Sets & Runs

Eval Sets: Combine personas + scenarios + evaluators as reusable test suites
Cartesian product: Total conversations = personas x scenarios x variants
Total evaluations: conversations x evaluators
Concurrency control: Max concurrent conversations configurable
Regression detection: Compare current run vs. baseline with configurable threshold (e.g., 0.1 = 10% delta)
CI/CD integration: API-triggered runs, regression checks block deployments

Result Interpretation

Pattern	Meaning	Action
High avg, low stdDev	Consistently good	Ship it
High avg, high stdDev	Good but unpredictable	Investigate outliers
Low avg, low stdDev	Consistently bad	Review instructions/flow
Low avg, high stdDev	Wildly varying	Check specific failing scenarios

Result Details

Per-evaluator score breakdown with standard deviation and confidence intervals
Individual conversation review: full transcript, scores with reasoning, trace, milestones
Pass@K metric for creative tasks
Regression detail: evaluator, persona/scenario, baseline vs. current, delta

16. 20+ Communication Channels

Deploy agents to any channel without changing agent logic:

Category	Channels
Messaging	WhatsApp, Slack, Microsoft Teams, Telegram, Messenger, Instagram, LINE, Twilio SMS
Voice	Jambonz (SIP), Twilio Voice, AudioCodes, LiveKit, BYOC SIP, Browser Voice, Kore Voice
Web	Web Chat Widget (SDK), WebSocket, AG-UI
Enterprise	Zendesk, Genesys, Email
API	HTTP, HTTP Async, Agent-to-Agent (A2A)

Channel-Specific Features

Channel	Rich Output	Threading	Media	Streaming
Web SDK	Full (HTML, Markdown)	N/A	Upload	SSE
Slack	Block Kit	Threaded	Files	No
Teams	Adaptive Cards	Threaded	Files	No
WhatsApp	Interactive (buttons, lists)	No	Media	No
Voice	SSML	N/A	N/A	Real-time
Email	HTML	Thread	Attach	No
Telegram	Keyboards	No	Media	No

Web SDK Setup

import { AgentSDK } from '@agent-platform/web-sdk';

const sdk = new AgentSDK({
  baseUrl: 'https://your-instance.ablplatform.com',
  projectId: 'proj_abc123',
  channelId: 'chan_xyz789',
});

await sdk.connect({ token: 'sdk_token_here' });
const chat = sdk.chat();
chat.send('Hello!');
chat.onMessage((msg) => console.log(msg.text));

Web SDK Features

Modes: Vanilla JavaScript, React hooks (AgentProvider, useAgent(), useChat(), useVoice()), web component
Chat: Message sending, attachment uploads (25MB max), typing indicators, rich content rendering
Voice: Voice interaction with VAD (voice activity detection), pipeline mode and realtime mode
Widget: Configurable position (bottom-right/left, top-right/left), theme customization (primaryColor, fontFamily)
Security: Domain restriction via allowedOrigins whitelist, user context passing (id, name, email, metadata)
Environment targeting: dev/staging/production

Slack Integration

Create Slack app with scopes: chat:write, app_mentions:read, im:read/write, files:read
Threaded replies to keep channels clean
Slash commands (e.g., /ask) routing to webhook
Interactivity: button clicks, menu selections
Multi-workspace support via generic webhook URL

WhatsApp Integration

Meta Cloud API or third-party (Infobip, Twilio)
Interactive messages: buttons, lists
Media support: images, documents, audio, video (auto-downloaded)

Voice Channels

Jambonz (primary SIP gateway), Twilio, BYOC SIP, AudioCodes VoiceAI Connect
STT providers: Deepgram, others
TTS providers: ElevenLabs, others
SSML support for fine-grained voice control
Browser-based voice via Twilio Client SDK

Rich Content Across Channels

Multi-channel templates: DEFAULT, MARKDOWN, HTML, VOICE, ADAPTIVE_CARD, SLACK, WHATSAPP
Carousels with images and action buttons
Interactive actions: buttons, select menus, input fields with form submission
SSML for fine-grained voice control
File attachments with OCR, transcription, and key frame extraction
Fallback to plain text RESPOND when channel doesn’t support rich content

17. Deployment & Operations

Agent Versioning

Version snapshots: Capture compiled ABL definition at a point in time
Version status: draft, testing, staged, active, deprecated
Version comparison: Side-by-side diff viewer showing additions/removals
Promotion workflow: Promote versions through status stages
Validation: Compilation errors rejected before version creation

Deployment Pipeline

Development -> Staging -> Production

Versioned manifests: Agent name -> version mapping (or “auto” for latest)
Entry agent: Typically the supervisor
Deployment lifecycle: active -> draining -> retired
- Active: Accepting new sessions
- Draining: No new sessions; existing sessions complete naturally
- Retired: Fully decommissioned
Rollback: Retire current, reactivate previous deployment
Model overrides: Per-agent model/temperature without changing ABL code
Auto-follow channels: Update automatically when new deployments are created
Only one active deployment per environment

Environment Management

Three built-in environments: dev, staging, production
Per-environment variables: referenced as {{env.KEY}}
Secret variables: write-only after saving (never exposed in API/logs)
Copy variables between environments
Feature flags via environment variables
Environment-specific model overrides and channel bindings

Git Integration

Supports GitHub (OAuth/PAT/GitHub App), GitLab, Bitbucket, generic Git (PAT/SSH)
Export format: agents/, tools/, project.yaml, project-lock.yaml
Push/pull ABL files between Studio and repository
Auto-sync with configurable conflict strategies: manual, local_wins, remote_wins
Auto-deploy on branch push (GitOps workflow)
Branch-per-environment mapping: develop->dev, staging->staging, main->production

Project Import/Export

Export preview: Manifest, lockfile, files, warnings with validation
Export format: JSON with manifest, lockfile, files map (folder or zip)
Import preview: Dry-run showing operations (create/update/delete)
Import validation: Applied with validation before changes take effect
CI/CD integration: Pull preview, apply if valid
Migration: Export from one workspace, import to another
Limits: 1 MB individual file, 50 MB total; max 1000 agents, 500 tools

18. Operations & Monitoring

Session Browser

Sortable, filterable session table with date range and agent filters
Full conversation transcripts
Agent conversation tree visualization (multi-agent branching)
Execution trace timeline with timing, payloads, decisions
Session export to CSV format

Trace System (Full Observability)

Every execution produces a span tree with these event types:

Event Type	What It Captures
`llm_call`	Model, tokens, latency, prompt/response
`tool_call`	Tool name, parameters, timing
`tool_result`	Return values, errors
`decision`	Routing decisions, condition evaluations
`handoff`	Agent transfers with context
`guardrail_eval`	Check results, actions taken
`state_change`	Variable assignments, memory updates
`error`	Error type, message, stack
`response`	Final response sent to user

Real-time trace streaming during test sessions
Trace analysis endpoint for automated diagnostics
Trace search: agent:name, status:outcome, model:name, duration:>2s, tokens:>5000
Filter by type, decision kind, span ID
Metrics: latency, token usage, tool success rate, guardrail block rate

Insights Dashboard

KPI cards: Sessions, Messages, Tokens, Estimated Cost, Containment Rate
Trend charts: Volume and containment over time
Cost breakdown: By agent, by model, by environment
Agent performance: Sessions, duration, resolution rate, messages per session, tokens, cost, error rate
Quality monitoring: Score trends, regression alerts
Customer insights: User behavior patterns, satisfaction signals
Voice analytics: Call duration, MOS scores, jitter, latency, barge-in rate, ASR score, TTS MOS, silence percentage, turns

Voice Analytics (Detailed)

Metric	Description
Session count	Total voice calls per period
Error count	Failed calls
Call duration	Average and distribution
MOS scores	Mean Opinion Score (inbound/outbound)
Jitter	Audio quality variance
Latency	Response time
Barge-in rate	User interruptions
DTMF fallback	Keypad-based fallback rate
ASR score	Speech recognition accuracy
TTS MOS	Text-to-speech quality
Silence %	Percentage of silence in calls
Turns	Average conversation turns

Human-in-the-Loop

Task inbox for approvals, data entry, review, decisions, escalations
SLA countdown and priority indicators
Transfer session monitoring across providers (SmartAssist, Genesys, NICE, Five9)
Human Task API:
- GET /api/projects/:projectId/human-tasks — View pending tasks
- POST .../human-tasks/:taskId/claim — Claim a task
- POST .../human-tasks/:taskId/resolve — Complete a task
- POST .../human-tasks/:taskId/assign — Reassign a task

Alerts

Alert rules: error rate spikes, volume changes, SLA breaches, eval score drops, deployment events, latency thresholds, token budget, abandonment rate
Notification channels: in-app, email, webhook

Custom Trace Dimensions

Extract business attributes from session data for segmentation and custom analytics views.

Natural Language Analytics

Query analytics data using natural language (e.g., “Show me error rate trends for the billing agent last week”).

19. API & SDKs

Authentication Methods

Method	Prefix/Header	Use Case	Lifetime
JWT Bearer Token	`Authorization: Bearer`	Studio sessions, user actions	Short-lived
API Key	`abl_` prefix	Server-to-server, CI/CD	Long-lived
SDK Session Token	`X-SDK-Token` header	Embedded widget sessions	Short-lived
Public API Key	`pk_` prefix	Client-side widget, origin-restricted	Long-lived

Conversation API

Send messages to agents, create/retrieve sessions, access conversation history
Streaming: SSE (Server-Sent Events) for real-time response streaming
Traces: Retrieve execution traces with detailed span trees
Sessions: Create, list, get by ID, send messages, get history

Management APIs

Agent Management: CRUD operations on agent definitions
Multi-Agent Workflows: Supervisor routing, handoff rules, delegation configuration
Tool Management: Create, configure, test tools (HTTP, code, MCP)
Deployment Management: Publish to environments, manage deployment lifecycle
Project Configuration: Create/update/delete projects, manage settings

Knowledge & Memory APIs

Memory API: GET/SET/DELETE operations on named memory stores
Contacts API: Create/read/update contacts, link to sessions, query with filters

Analytics APIs

Metrics aggregation by dimensions: category, event_type, agent_name, channel, hour, day
Available metrics: count, avg_duration, error_rate, p95_duration, sum_tokens, sum_cost
Event categories: session, message, llm, tool, agent, gather, flow, channel, deployment, search, voice, audit, evaluation, feedback, system
Session-level metrics, cost breakdown, session export
Voice analytics: Hourly aggregations with MOS, jitter, latency, barge-in, ASR, TTS metrics

Project Export/Import API

GET .../export/preview — Preview export metadata
GET .../export — Full export (manifest + lockfile + files)
POST .../import/preview — Dry-run import showing operations
POST .../import — Apply import with validation

Key API Features

SSE streaming for real-time responses
Offset-based pagination (limit/offset)
Rate limit headers (X-RateLimit-Remaining, Retry-After)
Standard error codes with clear resolution guidance
CORS support for widget embedding

ABL SDK (`@abl/core` + `@abl/compiler`)

Function	Purpose
`parse(text)`	Parse ABL text to AST
`parseAgent(text)`	Parse agent definition
`parseSupervisor(text)`	Parse supervisor definition
`parseAgentBasedABL(text)`	Parse agent-based format
`parseYamlABL(text)`	Parse YAML format
`validate(text)`	Validate ABL, return errors
`isValid(text)`	Boolean validity check
`serialize(ast)`	Convert AST back to ABL text
`compileABLtoIR(docs)`	Compile parsed docs to Intermediate Representation
`validateIR(ir)`	Validate compiled IR
`validateCrossAgentRefs()`	Validate cross-agent references
`validateFieldReferences()`	Validate field references
`parseCondition(text)`	Parse expression/condition
`expressionToPython(expr)`	Convert expression to Python
`tokenize(text)`	Lexer for custom tooling

20. Administration & Security

Workspace & Team Management

Resource Hierarchy: Organization (optional) > Workspace > Project Workspace Roles:

Role	Key Permissions
Owner	Full control, transfer ownership, delete workspace, manage billing
Admin	Manage members, models, secrets, connectors, view audit logs
Operator	Deploy agents, view analytics, manage env vars, view sessions
Member	Create/edit agents, run tests, manage project resources
Viewer	Read-only access

Project Roles: Admin (full project control), Developer (build/change project resources), Tester (verification, simulation, analytics), Viewer (read-only) Organization Roles: ORG_OWNER, ORG_ADMIN, ORG_MEMBER, ORG_BILLING Custom Roles: Tenant-scoped role definitions managed from workspace settings and used for project-level custom memberships. Canonical engineering detail lives in docs/features/workspace-sharing.md and docs/features/custom-project-roles.md. Member Management: Invite by email, pending invitation tracking (7-day expiry), resend/revoke, and role hierarchy enforcement.

Authentication

Email/password with optional MFA (TOTP with recovery codes)
Enterprise SSO:
- SAML 2.0: ACS URL, Entity ID, attribute mapping (email, firstName, lastName)
- OIDC: Discovery URL, Client ID, Client Secret, scopes
- Domain verification via DNS TXT record
- Force SSO enforcement with password fallback for Owner
MFA enforcement: Workspace-level with grace period (7/14/30 days)
JWT-based session tokens
API keys: Long-lived, prefixed abl_, scoped to projects, hash-based storage
SDK tokens: Short-lived, safe for browser
Channel authentication: HMAC signature, JWT, API keys

AI Model Configuration

Provider-Neutral Architecture — Register models at workspace level, assign to tiers:

Tier	Purpose	Example Models
Fast	Low-latency, high-throughput	Haiku, GPT-4o mini, Gemini Flash
Balanced	General-purpose	Sonnet, GPT-4o, Gemini Flash
Powerful	Complex reasoning	Opus, o3, Gemini Pro
Voice	Real-time voice interaction	GPT-4o Realtime, Gemini Live

Supported Providers:

Provider	Models
Anthropic	Claude 4 Opus, Sonnet, Haiku
OpenAI	GPT-4.1, GPT-4o, GPT-4o mini, o3
Azure OpenAI	GPT-4, GPT-4o (Azure deployments)
Google	Gemini 2.5 Pro, Flash; Gemini 2.0 Flash
Amazon Bedrock	Claude, Titan models
Custom/LiteLLM	100+ providers via unified interface

5-Level Model Resolution Chain (priority order):

Deployment override
Agent DSL declaration (EXECUTION.model)
Agent DB config (Studio settings)
Project DB config (tier mapping)
Tenant default

Operation Tier Overrides: Different tiers for extraction, validation, tool_selection, response_gen, summarization, reasoning, coordination, realtime_voice Fallback Chains: Automatic retry with alternative providers on failure Context Window Management: Tool result compression, prior turn truncation, conversation compaction (triggered by compaction_threshold) LLM Policies: Allowed providers, model allowlists, daily/monthly token budgets, rate limits, demo mode Credential Management: API key validation, workspace/personal scoping, active/inactive status, last used/validated timestamps

Security Posture

Area	Implementation
Encryption	TLS in transit, application-level encryption at rest, KMS for credentials
Tenant Isolation	Every query scoped to tenantId; cross-tenant returns 404
Audit Logging	Immutable logs for all admin actions, guardrail events, deployments
PII Protection	GATHER sensitive field masking, guardrail-based PII detection/redaction
Secret Management	Encrypted storage, never exposed in logs/traces/API responses; reference via `{{secret.KEY}}`
Rate Limiting	Tenant, project, user, and channel level limits
Input Validation	All API inputs validated/sanitized, payload size limits
Session Security	Project/tenant scoped, encrypted at rest, configurable timeouts, explicit termination

KMS — Bring Your Own Key (Enterprise)

Envelope encryption: Key Encryption Key (KEK) from your provider wraps short-lived Data Encryption Keys (DEKs)
Supported providers: AWS KMS (symmetric CMK), Azure Key Vault (RSA/AES), Google Cloud KMS (AES), External KMS (any REST-compatible)
DEK scope: Per project/environment
Key rotation: Automatic (if provider supports) or manual via new key version
Failure modes: Fail-closed (encryption/decryption fail) or graceful degradation (cached DEKs)

Secrets & Configuration

Secrets: Key-value pairs with encryption, workspace or project-scoped; rotation via generate new -> update -> revoke old
Environment Variables: Per project, per environment; reference as {{env.KEY}}; secret flag hides value in UI; bulk import via .env or CSV
Workspace Guardrails: Content safety categories with configurable thresholds (0.0-1.0) and actions (block/warn/log) at workspace level; layered with project-specific policies

Connectors & Integrations

Category	Connectors	Auth Types
Messaging	Slack, Teams, WhatsApp, Telegram	OAuth 2.0, API key
Communication	Twilio, SendGrid, SMTP	API key, basic auth
CRM	Salesforce, HubSpot, Zendesk	OAuth 2.0, API key
Webhooks	Inbound/outbound webhook support	HMAC, API key
Custom	Any REST API via HTTP connector	Configurable

Connectors have lifecycle states: active, expired (OAuth refresh needed), revoked. Schema discovery auto-detects available data fields.

Compliance

Framework	Status
SOC 2 Type II	Maintained (independently audited)
GDPR	Compliant (data minimization, right to erasure, data portability, DPA available)
ISO 27001	Aligned
CCPA	Compliant
HIPAA	Ready (Enterprise plan with BAA)

Data Protection

Data retention: Configurable per tenant, auto-purged
Right to erasure: Cascading deletion across all services
PII handling: Detection, tokenization, redaction in logs
LLM provider data: Varies by provider (Anthropic, OpenAI, Azure, AWS data policies)
Data residency: Region-specific deployments available
Responsible disclosure: security@ablplatform.com with 2-day acknowledgment, 5-day assessment SLA
Data Processing Agreement (DPA): Available on request

21. Billing, Plans & Limits

Plan Tiers

Feature	Starter	Professional	Enterprise
Members	5	25	Unlimited
Projects	3	20	Unlimited
Sessions/month	1,000	50,000	Unlimited
Tokens/month	1M	50M	Unlimited
LLM Providers	2	All	All
Knowledge Bases	1 KB, 100 docs, 10MB	20 KB, 10K docs, 100MB	Unlimited
Connectors	—	Yes	Yes
Guardrails	Basic	Full	Full + custom
SSO	—	—	SAML + OIDC
KMS (BYOK)	—	—	Yes
SLA	Best effort	99.9%	99.95% + custom
Support	Community	Standard (1 biz day)	Premium (1 hour critical)

Usage Tracking & Billing

Credit model: Allocated credits with burn rate tracking and low-credit alerts
Usage metrics: Sessions, messages, tokens, tool calls, estimated cost
LLM usage analytics: Daily trends by provider, model, agent
Billing periods: Monthly or annual with monthly tracking
Alerts: Credit low, usage threshold, health degraded, feature limit, SLA breach

Platform Limits

Category	Limits
Rate Limits	Auth routes: 20/15min; API routes: 100/min; Chat: configurable
Payload	10MB request, 1MB webhook, 25MB file upload, 100MB batch
Session	Max concurrent, configurable age/idle/service timeouts
Data Retention	Sessions: 30d, Messages: 90d, Traces: 30d, Audit: 90d (configurable)

22. Deployment Options

Option	Description
Cloud (SaaS)	Default. Sign up, open Studio, start building. Nothing to install or manage.
Self-Hosted (Enterprise)	On-premises deployment for data residency/compliance requirements. Same capabilities.

Self-Hosted Requirements

Requirement	Minimum (Dev)	Recommended (Production)
CPU	4 cores	16+ cores
Memory	16 GB RAM	64+ GB RAM
Storage	50 GB SSD	500+ GB NVMe
Runtime	Docker 24+, Node.js 20 LTS	Kubernetes 1.28+
Database	MongoDB 7.0+ (replica set)	MongoDB 7.0+ (replica set)
Cache	Redis 7+ (persistence, TLS)	Redis 7+ (persistence, TLS)
Analytics	Optional	ClickHouse

Self-Hosted Services

Service	Port	Purpose
Runtime	3112	Agent execution engine
Studio	5173	Browser IDE
Admin	3003	Administration console
SearchAI	3113	Knowledge pipeline
SearchAI Runtime	3114	Knowledge search
Docling	8080	Document processing (Python)
BGE-M3	8000	Embedding model (Python)
Preprocessing	8003	Text preprocessing (Python)

Self-Hosted Security Checklist

TLS on all endpoints
256-bit encryption key
32+ character JWT secret
Auth enabled on MongoDB/Redis
PII detection enabled
Rate limiting configured
IP allowlist
Audit logging enabled

Operational Features

Health checks: /health on Runtime/Admin, /api/health on Studio/SearchAI
Scaling: Stateless Runtime/SearchAI behind load balancer; session state in Redis; shared state in MongoDB
Zero-downtime upgrades: Rolling updates with maxSurge:1, maxUnavailable:0, graceful WebSocket drain
Vault providers: Environment vars, file, HashiCorp Vault, AWS Secrets Manager, Azure Key Vault, K8s Secrets
Database migrations: Auto-applied on startup; manual migrations available via CLI

23. Competitive Differentiation

Capability	Agent Platform 2.0	LangGraph	CrewAI	Bedrock Agents
Agent definition	Declarative DSL	Python code	Python code	Console + SDK
Per-step reasoning control	Yes	Manual	No	No
Multi-agent orchestration	Built-in (supervisor, handoff, fan-out, delegation)	Graph-based	Role-based	Single agent
Communication channels	20+ built-in	Custom	Custom	Limited
Visual editor	Studio (visual + code)	LangGraph Studio	No	Console
Knowledge/RAG	Integrated (SearchAI)	Custom	Custom	Knowledge bases
Guardrails	DSL-level (input + output + streaming)	Custom	Custom	Bedrock Guardrails
Evaluation framework	Built-in (personas, scenarios, LLM judges)	LangSmith	No	No
Multi-tenant isolation	Platform-level	Custom	Custom	AWS IAM
Voice support	Native (6+ voice channels)	Custom	No	No
Human-in-the-loop	Built-in (task inbox, approvals)	Custom	No	No
Deployment	Hosted or self-hosted	Self-hosted	Self-hosted	AWS only

24. Industry Examples

The platform includes production-quality agent templates across industries:

Industry	Example	Scale
Banking	Jupiter: 12-agent retail banking	Credit cards, loans, deposits, fraud, disputes
Airlines	SkyConnect: 6-agent travel service	Flight search, booking management, policy advisor
Retail	Face Frames: 5-agent optical store	Product discovery, order management, returns
Telecom	NOC Supervisor: network operations	Alarm triage, outage detection, vendor dispatch
Healthcare	Patient assistant	Appointment scheduling, symptoms, prescriptions
Travel	Hotel booking with parallel search	Flights + hotels + activities fan-out
Insurance	Claims processing	Intake, assessment, routing, approval workflows

Orchestration Pattern Templates

Pattern	Description
Sequential Pipeline	A -> B -> C chain processing
Router/Dispatcher	Central router to specialists
Hub-and-Spoke	Coordinator with satellite agents
Parallel Fan-out	Multiple agents in parallel, merge results
Tiered Escalation	L1 -> L2 -> Human with de-escalation
Human-in-the-Loop	Agent + approval gates
Specialist Swarm	Coordinator spawns specialist teams
Event-Driven Reactor	Agents respond to system events

Integration Patterns

Pattern	Description
API Gateway	Tool-based REST calls to backend services
OAuth 2.0 Connectors	Slack, Salesforce, HubSpot with auto-refresh tokens
Webhook Ingestion	Inbound events trigger agents via channel webhooks
Agent-to-Agent (A2A)	REST API or internal handoff protocol between agents
Async Approval Chains	Multi-step human approval workflows with webhook callbacks

25. Developer Experience

Getting Started (Under 5 Minutes)

Sign up at ablplatform.com (no installation required)
Create a project in Studio (or use AI-Guided Wizard)
Paste a 20-line ABL definition
Test in the built-in chat playground
Deploy to a web widget or any channel

Development Workflow

Tutorial Path

Tutorial	Topic	Duration	Level
1	Build your first agent	~30 min	Beginner
2	Build a scripted flow	~25 min	Beginner
3	Multi-agent & knowledge bases	~45 min	Intermediate
4	Safety, testing, publishing	~45 min	Intermediate

26. Community & Support

Getting Help

Channel	Description
Documentation	Comprehensive docs within Studio
Community Forum	Community forum for builders
Slack	Community Slack workspace
Stack Overflow	Tag: `abl-platform`
In-app Support	Built into Studio
GitHub Issues	Bug reporting with component labels
Public Roadmap	roadmap.ablplatform.com with priority voting

Support Tiers

Tier	Response Time	Availability
Community	Best effort	Forum/Slack
Standard	1 business day	Email/tickets
Premium	4 hours	Dedicated channel
Enterprise	1 hour (critical)	Dedicated + phone

Uptime Commitments

Plan	SLA
Free	Best effort
Starter	99.5%
Growth	99.9%
Enterprise	99.95%

Incident Response

Detection -> Acknowledgment (15 min) -> Investigation (30-min updates) -> Resolution -> Post-mortem (5 business days) Status monitoring: status.ablplatform.com with email/RSS subscriptions.

Partner Program

Early access to features, technical support, co-marketing, training materials, and partner directory listing.

27. Glossary of Key Terms

Term	Definition
ABL	Agent Blueprint Language — the enterprise control plane for agentic AI
Agent	Autonomous unit with goal, persona, tools, and rules
Supervisor	Coordination layer that routes conversations to specialist agents
Handoff	Context-preserving transfer of conversation to another agent
Delegate	Synchronous sub-agent invocation (call-and-return)
Escalate	Transfer to a human agent with full context
Fan-out	Parallel delegation to multiple agents, merged results
Flow	Step-based execution with deterministic transitions
Reasoning	LLM-driven autonomous decision-making (default mode)
Gather	Structured data collection from natural conversation
Guardrail	Safety policy enforced on input, output, or streaming content
Constraint	Business rule with enforcement action on violation
Session	Stateful conversation between user and agent(s)
Thread	Agent activation within a session (each agent gets its own thread)
Trace	Execution record with span tree of all events
Knowledge Base	Document collection with vector embeddings for semantic search
RAG	Retrieval-Augmented Generation — search + LLM
IR	Intermediate Representation — compiled ABL ready for execution
Channel	Communication endpoint (web, Slack, WhatsApp, voice, etc.)
Tenant	Isolated workspace with its own data, users, and configuration
Project	Container grouping agents, tools, knowledge bases, and settings
Model Tier	Abstraction layer: Fast, Balanced, Powerful, Voice
MCP	Model Context Protocol — standardized tool server interface
NLU	Natural Language Understanding — intent classification and entity extraction
SSML	Speech Synthesis Markup Language — voice output control

28. Summary: Why Agent Platform 2.0

Fastest path from idea to production agent — 20-line ABL definitions, browser IDE, AI-guided project creation, no infrastructure to manage
Unique per-step reasoning control — Mix deterministic and LLM-driven steps in one agent, controlling cost, latency, and predictability
Enterprise multi-agent orchestration — Supervisors, handoff, delegation, fan-out, escalation, and human-in-the-loop as first-class DSL constructs
20+ channels out of the box — Web, voice, WhatsApp, Slack, Teams, and more — agent logic stays the same
Integrated RAG pipeline — Upload documents, configure search strategies, connect live data sources — no custom retrieval code
Built-in evaluation framework — Personas, scenarios, LLM judges, trajectory evaluators, regression detection, CI/CD integration
Enterprise security as a foundation — SOC 2, GDPR, multi-tenant isolation, encryption, guardrails, audit logging, KMS/BYOK
Provider-neutral LLM support — Anthropic, OpenAI, Google, Azure, Bedrock, custom — with 5-level resolution cascade and automatic fallback chains
Full observability — Trace every LLM call, tool invocation, decision, and handoff with span-level timing and natural language analytics
Visual + code development — Teams collaborate with each member using their preferred editing mode, with AI Architect assistance
Complete lifecycle management — Version control, Git integration, environment management, GitOps workflows, import/export, rollback
Flexible deployment — Cloud SaaS or self-hosted with Kubernetes, Docker, and full operational tooling

Building Agents

Administration

Documentation Index

​Agent Platform 2.0: Comprehensive Overview

​1. Elevator Pitch

​2. Who Is This For?

​3. Platform Architecture

​Three Core Services

​End-to-End Flow

​4. ABL — The Agent Blueprint Language

​File Types

​Top-Level Sections

​Anatomy of an ABL Agent

​Execution Configuration

​One Agent, Two Execution Modes

​Per-Step Reasoning Control — The Differentiator

​Flow Control Constructs

​Interactive Actions in Flow

​Expressions & Built-in Functions

​Data Types

​Attachments

​5. Multi-Agent Orchestration

​Orchestration Patterns

​Session Hierarchy

​Supervisor Definition

​Supervisor Advanced Features

​Delegation (Call-and-Return)

​Fan-out (Parallel Delegation)

​Escalation to Humans

​Human Task System

​Project Orchestration Patterns (Pre-built Topologies)

​Real-World Scale: Jupiter Banking Example

​6. Studio — The Development Environment

​Key Capabilities

​AI Architect

​AI-Guided Project Creation Wizard

​Project Sidebar Navigation

​7. Tools & Integrations

​Tool Types

​Tool Declaration Syntax

​Tool Execution Hints

​Key Tool Features

​Code Sandbox Tools

​Tool Definition Example (HTTP)

​Error Handling

​8. Knowledge Bases & RAG

​Ingestion Pipeline

​Supported Formats

​Chunking Strategies

​Search Strategies

​Agent-Level Search Tools

​Search Configuration

​Live Data Connectors

​9. Data Collection with GATHER

​Key Features

​GATHER Example

​10. Memory & State Management

​Session Variables

​State Manipulation Directives

​Persistent Memory

​Auto-Recall (Remember/Recall Pattern)

​11. Safety & Guardrails

​Three-Tier Guardrail Implementation

​Input Guardrails

​Output Guardrails

​Streaming Guardrails

​Business Constraints

​Guardrail Providers

​Guardrail Policies (Centralized)

​Built-in Guardrail Templates

​Rate Limiting

​Audit Logging

​12. NLU & Multi-Language Support

​Intent Classification

​Embedding-Based Semantic Matching

​NLU Model Configuration

​Multi-Language Support

​13. Lifecycle & Hooks

​ON_START (Session Initialization)

​HOOKS (Lifecycle Event Handlers)