Definition
A task agent is an agent role optimized for implementation: it takes a scoped task and turns it into a small, reviewable change with validation and a build receipt.
Think: “construction crew,” not “survey crew.”
What good looks like
A task agent should:
- Work from explicit acceptance criteria
- Keep diffs small and local
- Add/adjust tests alongside code
- Run the gates it claims to pass (typecheck, unit tests, etc.)
- Produce evidence that crosses the trust boundary
Inputs a task agent needs
- A scoped problem statement (often from an explore agent)
- Constraints (versions, policies, perf budgets)
- A defined “done” state (tests passing, behavior verified)
If those aren’t present, the correct move is usually to stop and request exploration—not to guess.
How to avoid “big bang” diffs
- Prefer multiple small commits over one large change
- Add temporary instrumentation if it de-risks the change
- Treat formatting as its own commit
- Keep refactors separate from behavior changes
Common pitfalls
- Moving fast by skipping gates (creates future work)
- Mixing exploration and implementation (context thrash)
- Passing the wrong gate (running unit tests but not the integration lane)
- Introducing reward hacking (“tests pass” because you weakened the tests)
Pairing recommendation
Task agents work best paired with:
- An explore agent up front
- A critic pass via oppositional validation on the back end
Practical rule
If your diff is too big to review in five minutes, it’s too big.