Delegate

Use sub-agents when they can do focused work in parallel while the parent keeps architectural judgment, integration, and final verification.

Keep vs Delegate

Keep in the parent:

• Understanding the user's actual request and constraints
• Architecture, security, product, and release-risk decisions
• Cross-module integration
• Final review, test interpretation, and user-facing summary

Delegate to sub-agents:

• Read-only exploration over a bounded file set
• Mechanical edits with a clear file ownership boundary
• Focused test or lint runs
• Boilerplate generation from an explicit spec
• Independent checks that can run while parent work continues

Do not delegate tiny one-step tasks, ambiguous product decisions, destructive operations without a clear acceptance criterion, or final verification.

Launch Focused Work

Use agent for a focused child run. Launch independent children together so they can run in parallel.

Give each child a concrete outcome and acceptance criteria. Let it continue until the work is complete or it needs input. Do not invent time, turn or token budgets, or tell it to stop early because a percentage of a budget has elapsed. The runtime carries any limits explicitly configured by the owner.

Prefer provider-neutral model_strength over hardcoded model ids. Children inherit the active model by default (model_strength: "same"), including type: "explore", so pass model_strength: "faster" explicitly to get the cheaper same-family sibling for read-only exploration:

{
  "name": "config_audit",
  "prompt": "Inspect crates/tui/src/config.rs and crates/tui/src/settings.rs for duplicate model-default logic. Return file/line findings only; do not edit files.",
  "type": "explore",
  "model_strength": "faster",
  "cwd": "."
}

For code changes, give the child a precise write boundary and tell it not to revert unrelated edits. Keep implementation children capable with model_strength: "same":

{
  "name": "docs_patch",
  "prompt": "Update only docs/configuration.md to document the new [statusline] keys. Match the surrounding style. Do not edit other files.",
  "type": "implement",
  "model_strength": "same",
  "cwd": "."
}

Use fork_context: true only when the child genuinely needs the current conversation prefix. Leave it omitted for fresh, narrower context.

Route the Work

Choose the route per child from agent(action:"roster") (no provider request), not from memory. It lists each role's resolved route and the models you may pick.

• thinking overrides a saved tier on any role. Dial: off low medium high xhigh max ultra, plus inherit (role default, then session effort) and auto.
• model and model_strength apply to unpinned routes only. A pinned profile refuses a conflicting model; to choose a model, start an unpinned role such as general. model_strength: "faster" is the same-family faster sibling, "same" the session model.
• Read the receipt, not your request: child_route shows the provider, model and route source, and requested_reasoning vs effective_reasoning shows what the route actually ran.
• Route by stakes. Mechanical work: the cheapest sufficient route with thinking: "off" or "low". Implementation: a capable route at medium or above. Review and verification: a strong route at high, on a different route than the author when independence matters.

Evaluate and Verify

Sub-agent outputs are self-reports. Re-check material claims before relying on them:

• Read changed files directly.
• Run the relevant tests locally.
• Inspect unexpected diffs before committing.
• Verify externally visible or destructive claims against source data.

Prompt Shape

A good delegation prompt includes:

• The exact task
• Files or modules owned by the child
• Files or behavior the child must not touch
• Expected output format
• Acceptance criteria

Weak prompt:

Fix the settings bug.

Strong prompt:

Own only crates/tui/src/settings.rs and its tests. Preserve existing config key names. Verify with the existing checks that provider-specific API key changes do not restart provider onboarding. Return the changed paths and test command output.