loop-orchestrator

Overview

loop-orchestrator implements the Orchestrator-Workers pattern: it fans out N independent sub-tasks to parallel worker agents in a single wave, collects their intermediate artifacts, and returns a summary to the runtime. It never synthesizes the final deliverable — that is the spec-writer’s job.

When to use

Triage selects this archetype when a phase contains N independent sub-tasks — multi-source research, side-by-side comparisons, per-item analysis — where worker K’s output does not depend on worker K-1. For sequential pipelines where each step consumes the previous step’s output, use loop-plan-execute instead.

Inputs

A phase spec block from plan.yml:

phase: research
archetype: orchestrator-workers
workers:
  - id: w1
    model: sonnet
    prompt: "Research option A — scope, maturity, trade-offs."
  - id: w2
    model: sonnet
    prompt: "Research option B — scope, maturity, trade-offs."
artifact_dir: /tmp/loop-runs/<ts>/intermediates/research/
acceptance:
  - one .md per worker, 200-500 words
  - each must cite at least 2 sources
budget:
  max_total_tokens: 100000
  per_worker_timeout_s: 300

Outputs

Workers that fail are retried once with the failure mode in the prompt. Persistent failures are surfaced to the runtime — never silently dropped.

Example

# Triage selects orchestrator-workers for:
# "Compare Vite, Turbopack, and esbuild as build tools for our frontend"
#
# Dispatches 3 workers in parallel → 3 research artifacts
# spec-writer assembles the final comparison document in Phase 5