Lightweight agent coordination without a central orchestrator — a different approach #657

Been following this project with interest — the Kubernetes-native orchestration angle is exactly the right direction for teams that need enterprise-grade multi-agent workflows.

I've been exploring the other end of the spectrum: what does agent coordination look like with zero infrastructure? No central server, no orchestration layer — just a local daemon, plain JSON state files, and Slack as the coordination medium.

The result is claude-wg: each Claude Code session posts its plan to a private Slack channel, collects async feedback from collaborators, and gets a ✅ when they're satisfied. A local daemon listens via Socket Mode and handles the state. Conflict detection flags overlapping file changes across sessions before anyone commits.

The interesting design tension I kept running into: in a system without a central orchestrator, who decides when a plan is good enough to proceed? I ended up making approval explicit and directional — a collaborator reacts ✅ in Slack, or the owner approves from inside Claude. Either way it's a deliberate human signal, not an automated gate.

Curious whether the ambient-code approach has landed on a different answer to that question — and whether there's a world where something like this sits alongside a full platform as a lightweight on-ramp for smaller teams.