AutoGen
57,354Multi-agent conversation framework with programmable group chats.
Maturity tier details
Computed tier: 1A(raw score 7.4). Effective tier: 1A.
Facets:documented_in=0evaluated_on=0integration_surfaces=0.5paper_citations=3recent_commit=2stars=1.9
Source of truth:pipeline/maturity/score.py
Description
Microsoft Research open-source framework for building LLM applications as conversations between configurable agents (AssistantAgent, UserProxyAgent, GroupChat). Supervision primitives: per-agent system message scoping, conversable-agent turn control, code-execution config that gates whether an agent may run tools or code and in what environment, and GroupChatManager for bounded multi-agent discussion. v0.4 rewrote the core as an actor-model runtime (AgentChat on top of autogen-core).
Why it's included
Occupies prevention × pre_generation × fully_autonomous for the conversation-centric multi-agent archetype. Distinct locus from CrewAI: AutoGen scopes behavior via per-agent system messages and turn protocols rather than explicit role/task contracts, which is worth separating in the taxonomy.
Security notes
Open-source (MIT) from Microsoft Research. UserProxyAgent can execute code — configure code_execution_config to run inside a container (Docker) rather than on the host, or disable execution entirely for untrusted prompts. Check compatibility across the 0.2 → 0.4 rewrite when auditing existing integrations; the APIs are not identical.
Addresses failure modes
- Scope creep — Actions outside declared task boundary.
- Context pollution — Trajectory drift, hallucination spirals, overlong or off-task context.