A comprehensive reference of terms used throughout the Agent Platform 2.0 documentation. Terms are organized alphabetically for quick lookup.Documentation Index
Fetch the complete documentation index at: https://koreai-v2-agent-platform-dev.mintlify.app/llms.txt
Use this file to discover all available pages before exploring further.
A
ABL (Agent Blueprint Language)
The enterprise control plane for agentic AI — a schema-driven language purpose-built for multi-agent orchestration where deterministic governance meets autonomous reasoning. ABL spans the full control spectrum: delegate autonomously, supervise selectively, or lock down as a deterministic state machine. Agent definitions compile into immutable artifacts — every version is auditable, every deployment is reproducible, every change is governed. ABL is both human-writable and designed as a code-generation target for AI-authored blueprints. Related: ABL language overviewAgent
An autonomous unit of work defined in ABL that handles a specific domain or task. Each agent has a goal, persona, tools, and optional conversation flow. Agents can operate independently or collaborate through delegation and handoff patterns. Related: Agent concepts, Flow-based execution, Reasoning modeC
Channel
A deployment endpoint through which end users interact with your agents. Channels include web chat widgets, SDKs (JavaScript, React, iOS, Android), REST APIs, and third-party integrations (Slack, WhatsApp, Microsoft Teams). Each channel has its own configuration, authentication, and content format support. Related: Channels overviewCOMPLETE
A lifecycle construct in ABL that defines when an agent considers its task finished. EachCOMPLETE block specifies a WHEN condition, an optional closing response, and an optional STORE directive to persist results. When a completion condition evaluates to true, the agent ends the conversation or returns control to its parent.
Related: ESCALATE, HANDOFF
Connector
An integration that links an external data source to a knowledge base. Connectors handle ingestion from sources such as file uploads, web crawls, Google Drive, Confluence, SharePoint, and custom APIs. Each connector manages its own sync schedule and change detection. Related: Knowledge base, RAGConstraint
A business rule that must be satisfied before an agent proceeds with an action. Authors can group constraints under named labels such aspre_search or pre_booking, but those labels are organizational only. Runtime scoping comes from the rule itself, using constructs such as WHEN and structural BEFORE, plus ON_FAIL actions to respond, collect missing information, escalate, hand off, or redirect to a different step.
Related: Guardrail
D
DELEGATE
A multi-agent construct that allows an agent to dispatch a subtask to another agent and wait for the result. Delegation includes input/output mapping, a purpose description, timeout configuration, and failure handling. The delegating agent pauses execution until the sub-agent returns its result. Delegates can target local or remote agents via REST or A2A protocol. Related: HANDOFF, SupervisorE
Embedding
A numerical vector representation of text content used for semantic search. The platform generates embeddings during knowledge base ingestion, converting documents into vectors that can be compared for similarity. Embedding models (such as BGE-M3) run as part of the SearchAI pipeline. Related: Vector search, RAG, Knowledge baseESCALATE
A construct that transfers control from an AI agent to a human operator. Escalation triggers fire based on configurable conditions (such as user frustration, policy violations, or agent inability to resolve a request). Each trigger includes a priority level (low, medium, high, critical), contextual information for the human agent, and routing configuration. Post-resolution actions define what happens after the human completes the task.
Related: HANDOFF, COMPLETE
F
Feature gate
A mechanism that controls access to platform capabilities based on a tenant’s subscription plan tier. Feature gates determine which functionality is available (such as advanced guardrails, custom roles, SSO, or higher usage limits) and enforce entitlement boundaries at the API level. Related: Plan tierFLOW
An optional construct in ABL that adds structured, step-based execution to an agent. Each flow contains named steps with actions such as responding, calling tools, collecting user input (GATHER), branching on conditions, and transitioning between steps. Agents without a FLOW section operate in reasoning mode by default, using LLM inference to decide actions dynamically. Adding a FLOW does not change the agent’s type — it is still the same agent, with the option to enable or disable reasoning on individual steps viaREASONING: true/false.
Related: GATHER, Flow-based execution, Reasoning mode
Flow-based execution
An execution style where an agent includes aFLOW section in its ABL definition, providing structured, step-based conversation control. The agent moves through named steps with explicit transitions, branching conditions, and user input collection. Adding a flow does not create a different type of agent — it adds structure to the same agent. Within the flow, each step can set REASONING: true to use LLM-driven decision-making or REASONING: false for deterministic behavior, allowing fine-grained control over how each part of the conversation executes.
Related: Agent, Reasoning mode, FLOW
G
GATHER
A data collection construct within ABL that defines fields to collect from users during a conversation. Each field specifies a name, type, prompt, validation rules, and extraction strategy. GATHER supports LLM-based inference, pattern matching, hybrid extraction, field dependencies, progressive activation, and sensitive data handling (PII masking and transient storage). Related: FLOWGuardrail
A safety mechanism that evaluates agent inputs and outputs against defined policies. Guardrails canblock, warn, redact, escalate, fix, reask, or filter content based on rule-based checks, model-based safety classifiers (such as OpenAI Moderation or Azure Content Safety), or LLM-based evaluation prompts. Guardrails support streaming evaluation, graduated severity actions, and priority ordering.
Related: Constraint
H
HANDOFF
A multi-agent construct that transfers an active conversation from one agent to another. Unlike delegation, a handoff passes full conversational context (including variable state, conversation summary, and memory access) to the target agent. Handoffs can be one-way or round-trip (withRETURN: true), and support priority ordering, remote agent targeting, and asynchronous dispatch.
Related: DELEGATE, ESCALATE, Supervisor
K
Knowledge base
A managed collection of documents and data sources that an agent can search during conversations. Knowledge bases use retrieval-augmented generation (RAG) to provide agents with relevant context from your own content. Each knowledge base is scoped to a project and linked to one or more connectors that handle data ingestion, chunking, and indexing. Related: RAG, Embedding, Vector search, ConnectorM
MCP (Model Context Protocol)
An open protocol for connecting AI agents to external tools and data sources through a standardized interface. In ABL, tools can declare anMCP binding that connects to a configured MCP server, allowing agents to invoke external capabilities without custom HTTP integration code. The platform manages MCP server connections, authentication, and lifecycle.
Related: Tool, Sandbox
Member
See Role.Message
A single unit of communication within a conversation turn. Messages have a role (user, assistant, system, or tool) and content. The platform tracks messages as part of the conversation history, which is used for context management, compaction, and tracing.
Related: Turn, Session
MFA (Multi-factor authentication)
An authentication method that requires users to verify their identity through two or more factors before accessing the platform. MFA adds a layer of security beyond username and password, typically using time-based one-time passwords (TOTP) or authentication apps. Related: SSO, OIDC, SAMLO
Observatory
The platform’s real-time monitoring and debugging interface for agent sessions. Observatory provides visibility into active sessions, conversation history, LLM calls, tool invocations, state transitions, and trace events. Use Observatory to diagnose agent behavior, inspect decision paths, and replay sessions. Related: Trace, SessionOIDC (OpenID Connect)
An identity layer built on top of OAuth 2.0 that the platform supports for single sign-on authentication. OIDC enables your organization’s identity provider to authenticate users and issue identity tokens that the platform validates. Configure OIDC in your tenant’s SSO settings. Related: SSO, SAML, MFAOrganization
The top-level entity in the platform’s resource hierarchy. An organization groups one or more tenants under a single billing and administrative boundary. Organization members have roles such asORG_OWNER, ORG_ADMIN, ORG_MEMBER, and ORG_BILLING. Organizations are optional; tenants can operate independently without an organization parent.
Related: Tenant, Project, Workspace
P
Plan tier
A subscription level that determines the features, usage limits, and support options available to a tenant. Plan tiers control quotas for LLM tokens, concurrent sessions, knowledge base storage, and access to advanced capabilities. Each tier includes a set of entitlements that the platform enforces through feature gates. Related: Feature gateProject
A container within a tenant that groups related agents, knowledge bases, tools, and configuration. Projects provide resource isolation, access control, and deployment boundaries. Each project has its own members with role-based permissions, LLM configuration, environment variables, and SDK channels. Related: Tenant, Workspace, OrganizationR
RAG (Retrieval-augmented generation)
A technique that enriches LLM prompts with relevant content retrieved from knowledge bases before generating a response. RAG allows agents to answer questions using your organization’s data without fine-tuning the underlying model. The platform supports RAG, function-based, and hybrid retrieval strategies. Related: Knowledge base, Embedding, Vector searchReasoning mode
The default execution mode for every agent. In reasoning mode, an agent uses LLM inference to decide its actions dynamically based on its goal, persona, tools, and constraints, using strategies such as ReAct (reason + act), chain-of-thought, or direct inference. An agent without aFLOW section operates entirely in reasoning mode. Within a flow, individual steps can enable or disable reasoning via REASONING: true/false, allowing fine-grained control over which parts of the conversation use LLM-driven decision-making. Reasoning mode is not a separate agent type — it is the built-in behavior of every agent.
Related: Agent, Flow-based execution, FLOW
Role
A named set of permissions assigned to a user within the platform. Built-in roles include:| Role | Scope | Description |
|---|---|---|
| Owner | Tenant | Full administrative control, including billing and tenant deletion |
| Admin | Tenant | Manage members, projects, and configuration (cannot delete the tenant) |
| Member | Tenant | Create and manage projects assigned to them |
| Viewer | Tenant | Read-only access to tenant resources |
| Owner | Project | Full control over a specific project |
| Admin | Project | Manage project members, agents, and settings |
| Member | Project | Build and test agents within the project |
| Viewer | Project | Read-only access to project resources |
S
SAML (Security Assertion Markup Language)
An XML-based standard for exchanging authentication and authorization data between an identity provider and the platform. SAML enables enterprise single sign-on, allowing your employees to access the platform using their corporate credentials. Configure SAML in your tenant’s SSO settings. Related: SSO, OIDC, MFASandbox
An isolated execution environment for running user-defined code as a tool. Sandboxes support JavaScript and Python runtimes with configurable memory limits and timeouts. Code runs in a restricted environment with no access to the host filesystem or network, providing safe execution of custom logic within agent flows. Related: Tool, MCPSDK (Software Development Kit)
A client library for integrating the Agent Platform 2.0 into your application. SDKs handle session management, message transport, authentication, and UI rendering. Available SDKs include JavaScript (browser), React components, and REST API clients. Each SDK connects through a configured channel with its own API key. Related: Channel, WebhookSession
A stateful conversation between an end user and one or more agents. Each session maintains conversation history, variable state, active thread information, and gathered data. Sessions are scoped to a tenant and project, stored in Redis for active conversations and persisted to durable storage for cold sessions. Sessions support forking, thread stacks for multi-agent handoffs, and automatic expiration via configurable TTLs. Related: Turn, Message, ObservatorySSO (Single sign-on)
An authentication scheme that allows users to log in once with a single set of credentials and access the platform without re-authenticating. The platform supports SSO via SAML 2.0 and OpenID Connect (OIDC) protocols, enabling integration with enterprise identity providers such as Okta, Azure AD, and Google Workspace. Related: SAML, OIDC, MFASupervisor
A coordination layer defined in ABL that manages routing between multiple agents. The supervisor maintains shared state, defines routing rules based on intent matching and conditions, enforces policies, and controls communication settings. Supervisors determine which agent handles each user message based on priority-ordered routing rules. Related: Agent, HANDOFF, DELEGATET
Tenant
An isolated workspace within the platform that contains projects, members, and configuration. Every API request is scoped to a tenant, providing strict data isolation between different teams or customers. Each tenant has its own LLM policies, retention settings, and member roles. Tenants can optionally belong to an organization for centralized billing and administration. Related: Organization, Project, WorkspaceTool
An external capability that an agent can invoke during a conversation. Tools are defined in ABL with a name, description, parameters, and return type. The platform supports multiple execution types:| Type | Description |
|---|---|
| HTTP | REST API calls with configurable auth, headers, and retry |
| MCP | Calls to an MCP server via the Model Context Protocol |
| Lambda | Serverless function invocation (AWS Lambda or equivalent) |
| Sandbox | User-uploaded code running in an isolated runtime |