Show and Tell: ORCH — typed task queue + state machine for production AI agent teams

[oxgeneral]

Hey AgentOps community 👋

AgentOps covers the monitoring and observability side of AI agents — something I use too. I've been building the coordination layer: ORCH, a TypeScript CLI runtime for managing teams of AI coding agents.

The gap I was filling:
AgentOps tells you what an agent did. ORCH tells agents what to do next — and ensures they don't proceed without proper review.

Core primitive — typed task lifecycle:

type TaskStatus = 
  | 'todo'        // waiting in queue
  | 'in_progress' // claimed by an agent
  | 'review'      // agent completed, awaiting review
  | 'retrying'    // failed, being retried automatically  
  | 'done'        // reviewed and approved
  | 'failed';     // exhausted retry attempts

The review state is the key insight: no task reaches done without passing through a review gate. This catches a common failure mode in production agent systems — silent "success" that isn't actually validated.

Practical usage:

npm install -g @oxgeneral/orch
orch agent add "backend" --adapter claude --max-retries 3
orch task add "Implement payment API" --assignee backend --priority 1
orch tui   # TUI dashboard: states, PIDs, event streams

What would make the combination powerful: linking ORCH run IDs to AgentOps session IDs for end-to-end tracing from task assignment through agent execution.

Question: How are AgentOps users currently handling task assignment and retry logic? Is that typically custom application code, or are people using purpose-built orchestration tools?

GitHub: https://github.com/oxgeneral/ORCH (MIT, TypeScript strict)