Skip to content

Agents

An agent definition narrows a whole session to one job. Most day-to-day work never needs a direct agent invocation — commands route to the right skill on their own, and that's usually all you need. Invoke an agent directly when you specifically want that narrower framing, most commonly to force a fresh, unbiased pass over work someone else (or a different session) already did.

This page is the narrative layer. For the exhaustive, generated list of every agent this repo ships — full descriptions and category groupings — see the Agents reference. This page is about what agents are and when to reach for one directly; that page is the inventory.

How to invoke one

Syntax depends on your platform: @g-agnt-<name> on Cursor, your platform's native subagent or persona mechanism elsewhere — for example, Claude Code's Task tool with a matching subagent_type. See platform-usage.md for the per-platform pattern.

Why a fresh agent matters for verification

The clearest reason to invoke an agent directly instead of letting a command route for you: verification needs a different set of eyes. An agent that just implemented a change has already convinced itself the change is correct — that's not a criticism, it's just how self-review works everywhere, human or AI. g-agnt-verifier and g-agnt-code-reviewer exist specifically so review happens in a session that never saw the implementation reasoning, only the resulting diff and the acceptance criteria it's supposed to satisfy. This is also how autopilot's own implement/review pipeline works under the hood — a fresh reviewer agent, never the implementer, scores every pass.

Agents worth knowing

  • g-agnt-verifier — verifies a completed task against its acceptance criteria; never used by the agent that implemented the task.
  • g-agnt-code-reviewer — security, quality, and performance review with a structured, severity-rated report.
  • g-agnt-qa-engineer — bug reporting, tracking, and fix documentation; activates proactively on any error or defect mentioned in a session, even one that looks pre-existing or unrelated.
  • g-agnt-task-manager — task lifecycle: creation, status transitions, TASKS.md sync.
  • g-agnt-wpac-coordinator — triages incoming cross-project coordination items and routes them.
  • g-agnt-platformer — maintains cross-platform IDE/agent integration as each supported platform's own docs and capabilities change.
  • g-agnt-project-initializer — first-run project setup: scaffolding, workflow selection, initial skill install.

Full inventory

The Agents reference lists every shipped agent, grouped by category, with its full description. Use that page as your search target; use this one to get oriented first.