AutoGen is Microsoft Research's Python library for multi-agent conversation. Linchpin is an HTTP-API agent runtime with per-session Docker sandboxes. Library vs runtime — different layers of the same stack.
AutoGen is a Python library from Microsoft Research for building applications around conversations between multiple agents. You define agents (each one an LLM with a system prompt and tools), then orchestrate them in group chats, sequential conversations, or nested team patterns. AutoGen v0.4+ rewrote the core for async + actor-style messaging. License is MIT. The library is imported into your own Python service.
Linchpin is an open-source managed-agent runtime. Three services plus Postgres on a single VM. Agents, environments, sessions, events are HTTP resources. Each session runs in its own Docker container with controlled networking. The append-only event log streams over SSE with cursor replay. Language-agnostic — anything that can curl can drive Linchpin. Apache-2.0.
Frame: AutoGen is code you import into a Python process. Linchpin is infrastructure you deploy and call over HTTP.
| Dimension | Linchpin | AutoGen |
|---|---|---|
| What it is | Managed-agent runtime (platform) | Multi-agent conversation library |
| Maintainer | Open source community (linchpinhq) | Microsoft Research |
| License | Apache-2.0 | MIT (Creative Commons for docs) |
| Language | HTTP API; runtime in Python | Python library |
| Deployment | docker compose up on a VM | Imported into your Python app |
| Multi-agent | Run multiple sessions; orchestrate from your code | First-class group chat / sequential / nested patterns |
| Sandbox | Per-session Docker container, per-env network policy | Optional code-executor (Jupyter, Docker) you wire |
| Event log | Append-only Postgres, SSE replay | Conversation history in-process |
| Model providers | OpenRouter (~200) + Ollama | Broad (OpenAI, Azure, others via clients) |
| Streaming UI | SSE event stream out of the box | You build it |
| Persistence | Postgres by default — durable session state | In-process unless you wire your own |
They compose. AutoGen orchestrates the conversation pattern inside a Python service; Linchpin runs individual agents that need a sandbox (shell access, file ops, MCP tools). AutoGen calls Linchpin's HTTP API the way it would call any tool, and gets back streamed events from a real sandboxed session.
If your problem is "agents talking to each other" → AutoGen. If your problem is "agents executing tasks in isolated environments" → Linchpin. Many real apps are both.
AutoGen v0.4 is sophisticated and well-resourced. Linchpin is small and focused. They aim at different problems — multi-agent topology vs single-agent runtime infrastructure. Pick by which one matches your bottleneck.