Task-Shaped Agent Routing

Choose responsibilities and parallel work that improve the requested outcome. Model and reasoning settings come from the user's current routing policy, not from an assumed cost or capability ladder.

Respect the runtime contract

Honor an explicit model and reasoning preference for every role. In AX's Codex configuration the uniform route is gpt-6-astra with xhigh reasoning. Do not down-route scouts or workers, raise reviewer effort, or choose a different model merely because a role sounds cheaper or harder.

For other environments, inherit the caller's supported settings unless the user requests a different route. Do not silently translate model names into unsupported provider-specific fields. Distinguish requested configuration from observed runtime metadata, and disclose unavailable metadata.

Decide whether delegation helps

Keep work direct when it is tightly sequential, needs the coordinator's own source understanding, or is faster to perform and verify than to dispatch.

Delegate independent investigations, well-separated implementation areas, or one consequential claim needing independent evidence when parallel agents can save time or improve quality. Use parallel tool calls for independent lookups that do not need separate judgment.

Before splitting, consider source ownership, integration cost, observability, and the consequence of a wrong result. A large file count is not itself a reason to create agents. A small change may still benefit from an independent review of a concrete permission or persistence boundary.

Bound the task

Give each child the outcome, relevant sources, owned files, allowed mutations, expected evidence, and condition for returning. Add a retry limit when the operation can loop. Keep writers disjoint and coupled changes sequential.

Roles describe the work: retrieval, implementation, integration, diagnosis, research, architectural judgment, or focused review. They do not require separate model tiers or a fixed roster.

Never let a worker broaden permissions or perform an external action that the user has not authorized. Use actual tool capabilities rather than names copied from another harness.

Integrate and finish

The coordinator owns the stable objective, dependency order, and final acceptance decision. Read child evidence and check consequential claims. Incomplete or contradictory output calls for a better observation, a clarified contract, or direct takeover; it does not automatically justify another reviewer or a more expensive model.

Keep useful progress and uncertainty visible while continuing authorized work. Do not create a new task graph or ledger for work that fits one coherent pass.

Result

For a routing design, return the chosen direct/delegated structure, ownership, runtime settings, and verification owner. For implementation, apply that structure and deliver the requested result. Add independent review only for a named material claim; do not use agent count as a quality metric.