Test Agents

Test and debug Agentic projects before deploying them to production. The platform provides interactive test sessions, execution traces, session inspection, and automated evaluation workflows to help validate agent behavior, tool execution, routing, and response quality.

The platform supports the following testing capabilities:

Capability	Purpose
Studio Test Chat	Interactively test agents and workflows
Execution Traces	Inspect decisions, tool calls, handoffs, and state changes
Session Inspection	Analyze runtime behavior for a session
Trace Analysis	Detect slow spans, failures, and anomalies
Test Personas & Scenarios	Create repeatable conversation flows
Evaluations (Evals)	Measure quality and performance using automated evaluators

Test in Studio Chat

Use the Studio test chat to validate agent behavior interactively before deploying to production. The agent responds using the same execution pipeline as production, including:

• Tool calls and their results.
• Flow step transitions.
• Handoffs between agents.
• Guardrail checks on inputs and outputs.

Start a Test Session

1. Open your project in Studio.
2. To test the full project, go to the Overview page and click Test your agents. For multi-agent projects, Studio defaults to the project’s entry agent (usually the supervisor).
3. To test an individual agent in isolation, go to the Agent page and click Chat with Agent.
4. Or use the Chat option next to any agent on the Overview page.
5. Studio creates a test session. Type a message in the chat input and press Enter.
6. Each agent response message includes metadata such as:
   • Executed flow step
   • Tool invocations
   • Response time

Debug View

The Session Debug view provides real-time visibility into agent execution during a conversation. Use it to inspect responses, trace execution flow, monitor token usage, and diagnose runtime issues.

The left panel displays active and historical chat sessions. The center panel shows the live conversation between the user and the agent, including routed responses and workflow progression. The right-side diagnostics panel provides detailed runtime insights such as trace events, model usage, token consumption, session identifiers, and timeout diagnostics in real-time.

Signal	What it means	What to do
High response latency on a turn	A tool call is slow or the LLM loop ran multiple iterations	Check the trace for slow tool_call spans; consider adding ON_ERROR retry limits
Unexpected handoff	Supervisor routed to wrong agent	Review the handoff_match trace event and refine WHEN conditions
Tool called with wrong parameters	GATHER field mismatch or type error	Check gather_extraction trace events and validate field types in ABL
Guardrail triggered unexpectedly	Output hit a content or format boundary	Review guardrail_eval trace events and adjust guardrail thresholds
Agent asks for data already provided	Session state not passed during handoff	Check PASS declarations in HANDOFF blocks

Using Traces to Debug

Use traces to inspect every decision, tool call, state update, and handoff during execution.

1. Open your project in Studio.
2. Navigate to Sessions and select a session.
3. Click the Traces tab to view the full execution trace.
4. Use filters to narrow results by event type, agent, or time range.

Trace Structure

Each session produces a hierarchical execution trace organized into spans.

Session
+-- Turn 1
|   +-- LLM Call
|   +-- Tool Call
|   |   +-- Tool Result
|   +-- Response
+-- Turn 2
|   +-- Decision
|   +-- Handoff
|   +-- Response