Skill · pattern · subagents

Spawn subagents

Spawn a nested loop as a context boundary: written brief, fresh window, reduced tools, no parent secrets, structured return. Use when the parent would drown in search hits, specialists need isolation, or spawning researchers / reviewers.

GET /api/canon/skills/subagents?format=md

Claude Code, Deep Agents, and Codex all spawn subagents this way. Isolation that inherits every secret is theater.

When

Protect the parent window, parallelize, or apply a different skill set.

Do

01
Write the brief
Goal, constraints, return schema. The child does not inherit the transcript.
02
Reduce the principal
Fewer tools. No deploy keys. Read-only if the job is research.
03
Take a document back
notes.md plus a 30-line summary. Caps on depth, breadth, time.

Don't

Fork the parent process with every tool and token.
Return 40 search hits into the parent window.

Hard rules

Subagents exist to protect the parent window.
Depth-capped. Breadth-capped. Money-capped.

Refuse

Subagent inherits the parent — Nested loops get every tool and every secret. Isolation was theater.
Shared mega-context — A crew sharing one bloated window. Specialists drown in each other's work.

Load with this

Job · Research agent Job · Coding agent Playbook · Spawn subagents Subagents Small, Focused Agents Secrets, Identity, Least Privilege

Load next

Trigger tests

Should fire

“Spawn a researcher subagent”
“The parent context is drowning”

Should not

“Call a helper function in the same process”