I built OpenSwarm because I wanted an autonomous “AI dev team” that can actually plug into my real workflow instead of running toy tasks. OpenSwarm orchestrates multiple Claude Code CLI instances as agents to work on real Linear issues. It: • pulls issues from Linear and runs a Worker/Reviewer/Test/Documenter pipeline • uses LanceDB + multilingual-e5 embeddings for long‑term memory and context reuse • builds a simple code knowledge graph for impact analysis • exposes everything through a Discord bot (status, dispatch, scheduling, logs) • can auto‑iterate on existing PRs and monitor long‑running jobs Right now it’s powering my own solo dev workflow (trading infra, LLM tools, other projects). It’s still early, so there are rough edges and a lot of TODOs around safety, scaling, and better task decomposition. I’d love feedback on: • what feels missing for this to be useful to other teams • failure modes you’d be worried about in autonomous code agents • ideas for better memory/knowledge graph use in real‑world repos Repo: https://github.com/Intrect-io/OpenSwarm Happy to answer questions and hear brutal feedback.