STACK with AutoGen
Why this matters
AutoGen agents talk to each other. That's the framing. While LangGraph wires nodes and edges and CrewAI assigns roles to a Crew, AutoGen runs a GroupChat where every agent is a participant in a conversation, and the GroupChatManager picks who speaks next. Beyond conversation, AutoGen leans hard into code execution: an Agent can write Python, run it, and pass the output back into the chat. That combination of “argue and ship code” is what makes AutoGen distinct in the multi-agent space.
STACK plugs into the parts that need real-world reach. Functions you register with AutoGen agents become thin calls into STACK; STACK injects the credential and lands the action in your audit chain. The Python that AutoGen agents write at runtime can stay local for sandboxed compute, or it can invoke a STACK skill: a sealed-execution capability that runs in a Fly Machine, returns a typed result, and is audited end-to-end. Local-execution and sealed-execution as complements, not alternatives.
What it unlocks
- Per-speaker audit in a GroupChat. AutoGen's GroupChatManager picks a speaker each turn, and the speaker either says something or calls a function. STACK tags every credential retrieval, every outbound call, every detector fire with the speaker's agent ID. When a GroupChat goes off-script (an analysis agent decides to write to production instead of staging), the audit chain tells you exactly which agent in the conversation triggered which action, and at which turn.
- Code execution and sealed skills, side by side. AutoGen agents that write and execute Python locally are great for ad-hoc analysis and fast iteration. STACK skills are great for capabilities you want sealed, signed, and reproducible (a deploy step, a payment, a transformation that has to be verifiable). Mix them: let the agent prototype in local Python, then escalate to a sealed skill when the action needs to be production-grade. The conversation captures both, the audit chain captures both, the kill switch covers both.
- The actor model meets a centralized audit chain. AutoGen v0.4's actor model spreads agents across processes and machines for distributed multi-agent. STACK's audit chain is centralized regardless of how the agents are physically deployed: every action from every agent in the cluster lands in one operator-scoped, hash-chained log. Distributed agents, single source of truth.
Wiring it up
Install AutoGen alongside the STACK Python SDK, register the agents in your GroupChat, and wrap the functions you want to gate behind STACK. Step-by-step setup, working code, and how per-agent audit looks in practice:
Last reviewed 2026-05-08.