Show and Tell: ORCH — the team OS layer missing from Claude Code swarms (state machine + messaging)

[oxgeneral]

Hey ruflo community 👋

As someone who's been running Claude-based agent swarms, I hit a predictable scaling wall: individual agents are great, but coordinating multiple Claude Code instances gets messy without a proper runtime layer.

The failure modes I kept seeing:

• Agents completing tasks without a review gate → silent regressions
• One agent failing silently while others kept running → wasted compute + broken state
• No way for Agent A to tell Agent B "I changed the schema, adjust your approach"
• No single view to see what's actually happening across all agents

I built ORCH to solve these specific problems — a TypeScript CLI runtime for managing Claude-based agent teams:

State machine lifecycle:

todo → in_progress → review → done
             ↓
          retrying (on failure, configurable attempts)

Inter-agent messaging (the killer feature for swarm coordination):

orch msg send claude-backend "Schema updated to v3, recheck your models"
orch msg broadcast "All agents: staging environment is ready for testing"
orch msg inbox claude-reviewer   # check pending messages

Shared context store:

orch context set api-base-url "https://staging.example.com"
# Any agent in the team can read this

Live TUI dashboard — see all Claude instances, their current tasks, PIDs, and event streams in one terminal view.

1493 tests, TypeScript strict mode, MIT license.

Question for ruflo users: What's your current approach to agent-to-agent communication in your Claude swarms? Do you rely on shared files, or something else?

GitHub: https://github.com/oxgeneral/ORCH
npm: npm install -g @oxgeneral/orch