Agent
Read-only broad fan-out search. Use it for sweeps across many files, directories, or naming conventions where only the conclusion matters, not the file contents. Reads excerpts rather than whole files, so it locates code but does not review or audit it. State the breadth you need, e.g. "moderate" or "very thorough".
Agent
Implementation work that needs local judgment - feature work, bug fixes, design-sensitive refactors, integrations. The default executor for real development that is more than mechanical but doesn't need the frontier model. Give it the goal, the constraints, and the done-criteria; it owns the local design decisions…
Agent
Fresh-context adversarial verification of completed work. Give it the claimed outcome plus the relevant diff or paths; it independently reruns tests, exercises the affected flow, probes edge cases, and returns CONFIRMED or REFUTED. Read-and-run only; it never plans, edits, or fixes anything.
Agent
Fast read-only lookup agent. Use for any "where is X" or "how does X work" question that needs no judgment call: finding a file, a symbol, a config value, or a usage site. Cheapest way to gather facts before planning or editing; prefer it over reading files yourself when more than a file or two is involved.
Agent
Mechanical execution of fully-specified work - pattern refactors and renames, tests that follow an existing convention, documentation updates, bulk multi-file edits from an explicit spec, running a suite and fixing trivial failures. Use only when zero design decisions remain; give it the goal, the exact scope, and the…
Agent
Security-sensitive implementation after approval - authentication and authorization, secrets handling, cryptography usage, input validation, hardening, dependency remediation. Give it only an approved, stable contract to execute; all pre-approval analysis belongs to the security-review role instead.
Agent
Read-only security analysis before anything gets approved - authentication and authorization, secrets, cryptography, input validation, hardening, dependency risk, threat review. Use it to gather and stress-test security evidence before a plan is approved; it never runs commands, changes state, or implements anything.