3.2 — Test Agent-to-Agent Conversation

How Run Steps Reveal Delegation

When the orchestrator delegates to a specialist, the run on the orchestrator's thread produces a run step of type tool_calls. Each tool call entry contains:

• type: "connected_agent" (distinguishing it from function calls)
• connected_agent.name: the tool name of the specialist (e.g., billing_agent)
• connected_agent.output: the specialist's response text

You can retrieve these steps programmatically:

run_steps = project_client.agents.list_run_steps(
    thread_id=thread.id,
    run_id=run.id,
)
for step in run_steps:
    if step.type == "tool_calls":
        for tc in step.step_details.tool_calls:
            print(f"Agent called: {tc.connected_agent.name}")
            print(f"Specialist output: {tc.connected_agent.output}")

Common Multi-Agent Issues and Fixes

Crafting Test Queries That Exercise Routing

Good test queries for multi-agent systems should:

• Target a single specialist — verify each specialist is reachable individually.
• Span multiple specialists — verify the orchestrator can make two delegation calls in one turn.
• Be ambiguous — verify the orchestrator routes to the correct specialist even when the query does not name the domain explicitly.
• Include edge cases — queries that sit on the boundary between two specialists' scopes stress-test description quality.

Example test matrix for a three-specialist orchestrator (billing, HR, tech support):

Validating End-to-End Response Quality

Multi-agent responses can fail at the seam between the specialist's output and the orchestrator's synthesis. Check for:

• Factual consistency — does the orchestrator's final reply accurately reflect the specialist's output, or did it hallucinate a summary?
• Completeness — did the orchestrator include all parts of the specialist's answer, or silently drop details?
• Attribution — if the specialist cited sources or returned structured data, does the final response preserve them?
• Latency — each additional agent hop adds ~1–3 seconds; validate that total response time is within UX requirements.

Debugging Multi-Agent Flows Step by Step

1. Check the orchestrator run status — a status of failed or expired signals a problem before even reaching the specialist.
2. List run steps — look for tool_calls steps. Their presence confirms delegation occurred.
3. Inspect the tool call input — the orchestrator passes a message to the specialist; verify it contains the right context.
4. Inspect the tool call output — verify the specialist's reply is accurate and complete.
5. Check the specialist's own thread — if the specialist has a nested thread, retrieve it to see its own run steps and any sub-tool calls.
6. Review agent call logs — in the Foundry portal, navigate to Monitoring → Traces to see latency and error detail for each agent hop.