← all projects
dAIlog — Multi-agent framework, Jan 2024
GPT-3.5 was new. Claude Code did not exist. OpenClaw did not exist. I had seven agents talking to each other.
The Python service spun up a real GroupChat: a user prompt would trigger Planner → Engineer (writes code) → Scientist (no code, classifies/reads) → Critic (doublechecks claims and source URLs) → Executor (runs the code) → loop until termination. Each agent had a real system prompt; the whole thing ran on gpt-3.5-turbo and gpt-4 with autogen.config_list_from_json. The C# layer handled auth, history persistence, and the chat UI. I proposed it as an internal product. It went nowhere internally. Six months later the same primitive was everywhere.