Sub Agents

Sub-agents

Read only the workflow needed for the task:

• PR, diff, or plan review: review.md
• Coding: implementation.md
• Diagnosis, research, or design exploration: investigation.md

Repository instructions may add project-specific guidance. When a repository defines its own sub-agent workflow, that guidance wins; this skill is the default everywhere else.

Model and harness selection

Select the model before starting any sub-agent, agent team, skill, or workflow:

1. Use the model the user requested.
2. If the user did not name one, apply the defaults below.
3. Choose a harness that supports that exact model.

Do not substitute a model to fit the harness already open. A model being unavailable inside one workflow is a harness limitation, not general unavailability. In Claude Code, Claude's native Agent and Workflow tools run Claude-family models; use codex exec from the start for GPT models such as gpt-5.6-sol. Do not start a bundled code-review Workflow and add the selected GPT reviewer afterward. In Codex, use the Claude CLI when the selected model is only available there.

Defaults

• Delegation is optional. Do straightforward work directly.
• One sub-agent may cover several relevant roles when that is the best use of an expensive model. Treat its conclusions as one perspective, not several independent opinions.
• Default to one sub-agent for the initial review. If accepted findings lead to code changes, the terminal round uses the original reviewer and the independent scope judge defined in review.md; this is part of the review cycle. Any broader initial coverage or additional reviewers still require the user's approval unless already requested.
• The primary agent owns the result. Verify claims, remove duplicates, and resolve disagreement before reporting or integrating work.
• Do not show one reviewer another reviewer's conclusions before both finish.
• Before starting more than one sub-agent for any purpose — review, investigation, benchmark, implementation — state how many you will run, the model each uses, and the rough cost, then wait for approval. A number the user gave ("two Opus 5") is the number; never expand a request into a larger matrix.
• Before a parallel fan-out, run the machine-resources check from the global AGENTS.md (load, memory, temperature). On a strained machine, run agents sequentially or in the cloud.

Models

• Default an unqualified PR, diff, or plan review to gpt-5.6-sol with high reasoning.
• Default other difficult GPT work to gpt-5.6-sol with high reasoning.
• Reserve gpt-5.6-sol with max reasoning for unusually difficult reviews or investigations.
• Opus 5 with high reasoning is a good general sub-agent for investigation, implementation, and review.
• Use Fable 5 only with high reasoning, mainly for review, investigation, or an independent perspective.
• Sonnet 5, gpt-5.6-terra, and gpt-5.6-luna are available but are not defaults. Use them when a bounded task has a specific speed, cost, or capacity reason.
• Do not use Haiku.
• Economics: when one provider is materially cheaper for you, prefer it where two models fit equally. User-facing or design-sensitive work still goes to a high-taste model (Opus 5 or Fable 5) regardless of cost.
• Gemini runs through the local agy CLI: agy --model gemini-3.7-flash-high --sandbox --dangerously-skip-permissions --print-timeout 10m --print "<prompt>". Use it for video, audio, and other multimodal input; scraping YouTube through a browser is throttled and does not work.

Context

Give a sub-agent enough context to work independently. When context is not inherited, include:

• the objective;
• repository and relevant paths;
• PR, issue, base branch, or diff scope;
• constraints that change the answer;
• the expected output;
• an instruction to read the repository's AGENTS.md and the docs/guidelines/ files matching the task's area before starting.

Do not add background merely because it exists. Reviewers should not receive the primary agent's conclusion unless their job is to challenge that exact conclusion.

Harnesses

Prefer the native sub-agent tool when it supports the requested model and isolation. Use the CLIs when the model is unavailable natively or the task needs a separate harness.

Claude Code, read-only:

claude --model <model> --effort high --permission-mode plan -p "<prompt>"

Claude Code, implementation:

claude --model <model> --effort high --permission-mode acceptEdits \
  --allowedTools "Bash(<repository lint and typecheck command>),Bash(<repository test command>)" \
  -p "<prompt>"

acceptEdits does not approve every Bash command. For a headless implementation, allow only the test and verification command patterns the task needs; do not allow unrestricted Bash merely to avoid prompts.

Codex, read-only:

codex exec -m <model> -c 'model_reasoning_effort="high"' -s read-only -C <repo> "<prompt>"

Codex, implementation:

codex exec -m <model> -c 'model_reasoning_effort="high"' -s workspace-write -C <repo> "<prompt>"

Parallel reviewers may share a checkout because they are read-only. Parallel implementation agents need separate worktrees and non-overlapping ownership.