ORCH — built an orchestration layer for teams of LLM-powered agents

[oxgeneral]

Hey Simon and community 👋

Long-time llm user here. I built ORCH — a CLI runtime that sits one layer above tools like llm, treating your AI tools as a coordinated engineering team rather than individual commands.

The mental model

llm is excellent for running a single AI command. ORCH answers: what happens when you have 5+ llm calls that need to be coordinated, retried, sequenced, and audited?

Think of it as: llm is a worker, ORCH is the team manager.

What ORCH adds on top

• State machine — tasks flow todo → in_progress → review → done with mandatory review gate before completion
• Auto-retry — failed agents automatically enter retrying state with configurable attempts
• Scope locking — prevents two agents from touching the same files simultaneously
• Inter-agent messaging — agents broadcast results to each other (orch msg send, orch msg broadcast)
• Shared context store — agents read/write a shared key-value store across runs
• Multi-adapter — orchestrates Claude Code, OpenCode (Codex/Gemini), Cursor, and plain shell scripts from a single CLI
• TUI dashboard — live Ink/React dashboard showing task queue, agent status, event feed

Quick example

npm install -g @oxgeneral/orch

# Define your agents
orch agent add "Researcher" --adapter shell --role "Run llm queries and summarize results"
orch agent add "Writer" --adapter claude --role "Turn research into documentation"

# Create tasks with dependencies
orch task add "Research async patterns" --assignee researcher
orch task add "Write docs" --depends-on tsk_abc123 --assignee writer

# Launch
orch run
orch tui  # live dashboard

The state machine means the Writer agent only picks up the writing task after Research has passed review — no polling, no manual coordination.

Why I built it

I was running multiple llm instances manually for a research pipeline — copy-pasting outputs, restarting failed runs by hand. ORCH makes that workflow reliable and observable.

Links

• GitHub: https://github.com/oxgeneral/ORCH
• npm: npm install -g @oxgeneral/orch
• 1,657 tests, MIT license, TypeScript strict

Curious if others here have built multi-step pipelines with llm — would love to hear how you handle coordination today. Happy to answer any questions!