Skill · pattern · orchestrator-workers
Orchestrator–workers
A parent decomposes work at runtime, spawns isolated workers, and synthesizes. Use for long-horizon research, DeerFlow/Deep Agents shape, or when subtasks are discovered during the run rather than at design time.
GET /api/canon/skills/orchestrator-workers?format=md
Anthropic catalog + Deep Agents / DeerFlow. Workers are context boundaries. Returns are documents, not transcripts.
When
Subtasks are not known up front. The parent must not drown in search hits.
Do
- 01
Parent plans and synthesizes
It does not read 40 hits. It writes briefs and reads 30-line summaries plus a notes path.
- 02
Workers get a fresh window
Reduced tools. No parent secrets. Caps on depth, breadth, time.
- 03
Notes on disk
Filesystem is source of truth. The window is a cache.
Don't
- Dump worker transcripts into the parent.
- Let workers spawn unbounded children.
Hard rules
- A second agent is a context boundary, not a personality.
- Return schema is part of the brief.
Worker brief
python
brief = {
"goal": "Find prior art for feature flags in this repo.",
"constraints": ["read-only", "no network", "10 minute cap"],
"tools": ["grep", "read_file", "glob"],
"return_schema": ["notes_path", "summary_30_lines", "open_questions"],
}
result = spawn_subagent(brief=brief, context="fresh", secrets=[], budget=Budget(max_steps=12))
Refuse
- Shared mega-context — A crew sharing one bloated window. Specialists drown in each other's work.
- Subagent inherits the parent — Nested loops get every tool and every secret. Isolation was theater.
- Context hoarding — Raw HTML, full CI logs, six PDFs, entire JSON envelopes stuffed 'just in case'. The goal is the first thing forgotten.
Load with this
Load next
- Spawn subagentsProtect the parent window, parallelize, or apply a different skill set.
- Research agentThe path is unknown, sources are many, and the deliverable is a brief.
- Fan-out / reduceIndependence is real. Shared mutable state makes this a race, not a pattern.
- Design a harnessThe product is a desk, not a chat wrapper. Claude Code, Codex, OpenHands, Cline, Goose, Deep Agents converge here.
Trigger tests
Should fire
- “Long-horizon research with subagents”
- “DeerFlow-style orchestrator”
Should not
- “Named three-stage pipeline”