ORCH — CLI-native agent team runtime (state machine, messaging, retries for coding agent workflows)

[oxgeneral]

What I built

ORCH — a TypeScript CLI runtime for teams of AI coding agents.

For agno/phidata users building multi-agent systems

Agno is excellent for building agent workflows in Python. ORCH fills a different gap: CLI-native orchestration for coding agents (Claude Code, Cursor, OpenCode) that run as OS processes.

The key problem I built ORCH to solve: when you run multiple AI coding agents in parallel, there's no accountability layer — agents mark themselves "done" without anyone verifying the output.

ORCH's state machine approach

todo → in_progress → review → done
                    ↑
               retrying (auto on failure)

Tasks require explicit review before completion. Agents communicate via typed messages:

orch msg send agt_reviewer "Feature X ready for review"
orch msg broadcast "Shared context: API changed to v2"

Shared context store lets all agents see the same data:

orch context set api-version "v2.1"
orch context get api-version  # available to all agents

Stats

• 1647 tests, TypeScript strict, MIT
• npm install -g @oxgeneral/orch

Repo: https://github.com/oxgeneral/ORCH