Subagent Delegation

Subagent Delegation

Treat the main session as the orchestrator. Keep planning, user judgment, synthesis, and final drafting in the main conversation. Delegate bounded gathering, reading, checking, and repetitive work when doing so saves time, cost, or main-context space.

Do not assume that a named agent, model, or reasoning level exists. ChatGPT Work uses hosted agents. Local Codex can also load custom TOML agents. The available catalog can differ by product, account, client, and release.

Start with the active runtime

Before spawning anything, inspect the collaboration or subagent tool schema shown in the session. Treat a role, model, and effort combination as available only when that schema advertises it.

Apply this order:

1. Use org-researcher, org-summarizer, org-bulk-worker, org-fact-checker, or org-verifier when the runtime advertises those installed agents. Use terra-ingest for long-context ingestion and sol-reviewer for an adversarial read-only review when those installed agents are available.
2. Choose the specialist by task, not by model preference:
   • org-researcher is the default for multi-source and source-constrained research.
   • org-summarizer is the default for faithful compression of long notes, documents, threads, or transcripts.
   • org-bulk-worker is the default Luna route for high-volume, bounded, low-judgment work.
   • org-fact-checker is for a claim-by-claim check of a draft against sources it opens itself; read-only, returns a verdict, not a rewrite.
   • terra-ingest is for large source sets or broad file/repository sweeps that need sustained context.
   • org-verifier and sol-reviewer are for independent, read-only checks.
3. Otherwise choose the smallest advertised built-in role that fits the work. Common examples include:
   • luna-leaf: one small, sharply bounded lookup or mechanical leaf task. In runtimes where it is fixed to Luna/xhigh, do not describe it as configurable Luna/high, medium, or low.
   • terra-ingest: long-context reading, many files, or a source sweep that must remain coherent.
   • sol-reviewer: independent, read-only verification when the stakes justify a stronger second pass.
   • ops-executor: three or more integration operations when that role is advertised and the user has authorized those operations.
4. Use an explicit model or reasoning override only when the active tool schema lists that exact value. Never invent a model slug or claim an effort level the runtime did not provide.
5. If no suitable lower-cost worker is advertised, work inline. Do not spawn an inherited or unknown-cost agent merely to say the task was delegated.
6. If no subagent capability is available, work inline and never imply that a worker ran.

When delegation pays

Delegate when at least one condition applies:

• Two or more independent, bounded work units can run in parallel.
• The task needs at least three searches, document reads, or integration calls.
• Intermediate source material, logs, or tool output would pollute the main context and can be compressed into a short return contract.
• An independent read-only check materially reduces the risk of an important send, published number, or difficult-to-reverse decision.

Handle the task inline when it is one quick lookup, requires the user's judgment at several points, or edits text already in the conversation. Keep emails, messages, public-facing prose, and other voice work in the main context. Spawning has fixed overhead and every worker consumes tokens.

Assign work without duplication

Give each worker one bounded task and exclusive ownership of its sources, paths, records, or questions. Include all context the worker needs because a worker may start blank:

• objective and definition of done;
• exact sources, paths, records, or questions it owns;
• sources, paths, and actions it must not touch;
• whether the task is read-only or permits edits or external actions;
• relevant names, dates, org rules, and user constraints;
• exact return contract.

Do not repeat a worker's searches in the main session while waiting. Spot-check the returned evidence or rerun only the narrow query needed for verification. Assigning the same item twice is allowed only when the second assignment is an explicit read-only verification pass.

Keep delegation to two layers. Allow a worker to fan out only when that middle layer genuinely compresses several sources into one result and the runtime permits the additional depth.

Return contracts

Use the contract that matches the job.

Researcher

• SUMMARY: 3-5 usable bullets.
• FINDINGS: each claim with source name, link or path, and date; show conflicting evidence both ways.
• UNVERIFIED: anything not confirmed.

Summarizer

• HEADLINE: one sentence.
• KEY POINTS: 3-7 concrete bullets.
• DECISIONS / ACTIONS: who, what, and by when, only when present.
• OMITTED: what was excluded and where it remains available.

Bulk worker

• DONE: count processed.
• SKIPPED: each skipped item and reason.
• SAMPLE: three processed results to spot-check.
• Stop instead of improvising when more than 10 percent of items do not fit.

Fact checker

• VERDICT: clear to send, fix first, or blocked.
• WRONG: claim as written, what the source says, source and date.
• UNVERIFIED: claim, where you looked, what would settle it.
• RIGHT BUT MISLABELED: value correct, framing wrong.
• CHECKED AND CORRECT: a count.
• Read-only: reports the gap, never rewrites the draft.

Verifier

• VERDICT: pass, fail, or pass-with-issues.
• ISSUES: exact location, problem, and evidence.
• UNVERIFIED: claims that could not be checked.
• Remain read-only. Do not edit, send, post, or fix.

Local Codex custom agents

Local Codex recognizes standalone TOML agents in ~/.codex/agents/ for one user or .codex/agents/ for a trusted project. Hosted ChatGPT Work does not have a documented mechanism for importing or persisting these local TOML files.

This skill includes nine Codex templates under assets/codex-agents/ and a safe, non-overwriting installer at scripts/install-codex-agents.sh. Run the installer only when the user explicitly asks to install the local agents. The templates put bounded research, summarization, and bulk work on Luna; reserve Terra for sustained-context ingestion, and Sol for verification, fact-checking, or difficult independent reasoning. They remain separate from the runtime-aware Work-mode fallback above.

Verification

Re-open changed files, rerun the relevant test or query, and spot-check worker claims before reporting completion. Use a read-only verifier when stakes justify the extra cost. If no suitable verifier is available, perform a narrow inline second pass and label remaining uncertainty.