Skip to content

autogen_adapter.md

Latest commit

 

History

History
68 lines (48 loc) · 2.12 KB

autogen_adapter.md

File metadata and controls

68 lines (48 loc) · 2.12 KB
description Wrap an AutoGen team as a TraceCore agent

AutoGen adapter quickstart

This tutorial shows how to generate a TraceCore-compatible agent from a Microsoft AutoGen team. The goal is to make framework-based agents measurable, repeatable, and testable inside TraceCore's deterministic harness.

Use this when you want to:

  • Keep your AutoGen team definition intact.
  • Plug it into TraceCore tasks with minimal glue.
  • Prove that agent behavior can be regression-tested like a unit test.

Requirements

Install the AutoGen dependencies used by the generated agent:

pip install autogen-agentchat autogen-ext[openai]

You also need tracecore installed (or a local editable install).

Generate an AutoGen-backed agent

Create a generator script:

from agent_bench.integrations.autogen_adapter import generate_agent

generate_agent(
    task_ref="rate_limited_api@1",
    model="gpt-4o-mini",
    agents=[
        {"name": "Worker", "system_message": "Execute tools precisely. Output one JSON action then say DONE."},
        {"name": "Supervisor", "system_message": "Review the action. Correct if wrong. Say DONE."},
    ],
    output_path="agents/autogen_rate_limit_agent.py",
)

Run it, then execute the agent:

python path/to/generate_autogen_agent.py
agent-bench run --agent agents/autogen_rate_limit_agent.py --task rate_limited_api@1 --seed 42

Determinism checklist (pytest-for-agents mode)

The generated agent includes a deterministic state machine and an LLM fallback. For reproducible runs in CI, use the adapter like this:

  1. Solve the task in the deterministic rules first.
  2. Gate or remove the LLM fallback in act() for CI runs.
  3. Record a baseline with agent-bench baseline --agent ... --task ... --export latest.
  4. Compare future runs with agent-bench baseline --compare run_a run_b.

This keeps AutoGen teams compatible with TraceCore's "proof of behavior" artifact flow while still allowing you to iterate on LLM behavior locally.

Where the adapter lives

  • Adapter source: agent_bench/integrations/autogen_adapter.py
  • Integration notes: agent_bench/integrations/README.md