Agents Reference

Claude Kit ships 8 specialist agents. Each agent has a single dispatcher (the skill that calls it) and a narrow job. No agent-bloat; no orphans.

How Agents Work

When a skill needs deeper, focused work, it dispatches a specialist agent. The agent starts in a fresh context, does the focused job, and returns a structured result to the main conversation.

You: "Review this code for security issues"

→ /claudekit:code-review-loop dispatches
  → claudekit:security-auditor (sensitive path detected)
  → Focused OWASP-aligned review
  → Returns findings with severity + OWASP category

Agents can be dispatched in parallel — plan-review runs architect and experience-reviewer simultaneously.

The 8 agents

Agent	Job	Dispatched by
claudekit:planner	Decompose specs into executable plans (file paths, exact test commands, acceptance criteria, Risks section)	`write-plan`
claudekit:architect	Score architecture dimension of a written plan: data flow, failure modes, edge cases, test matrix, rollback safety	`plan-review-architecture` (via `plan-review`)
claudekit:experience-reviewer	Score UX + DX dimension: information hierarchy, state coverage, accessibility, DX ergonomics, AI-slop avoidance	`plan-review-experience` (via `plan-review`)
claudekit:investigator	Root-cause investigation with evidence chain — never guesses, never patches symptoms	`investigate-root-cause`, `evidence-driven-debugging`
claudekit:tester	Design and write tests with red-green discipline; pastes runner output as evidence	`test-first`
claudekit:code-reviewer	Pre-merge structural review of diffs: error handling, edge cases, complexity, naming. Defers sensitive paths to security-auditor	`code-review-loop`
claudekit:security-auditor	OWASP-aligned review of sensitive paths (auth, payments, crypto, sessions, tokens)	`code-review-loop` (sensitive paths only)
claudekit:scout	Codebase mapping and dependency audits — produces evidence-cited maps with `<file:line>` references for every claim	`map-codebase`, `audit-dependencies`

Custom agents

You can add project-specific agents in .claude/agents/. They follow the same YAML frontmatter format as bundled agents:

---
name: my-agent
description: "When to dispatch this agent..."
tools: Read, Edit, Bash
memory: project
---

You are a [role] who [does what]. Your output is...

Agent design rules:

One dispatcher per agent. No orphans. If you can’t name the skill that dispatches the agent, the agent shouldn’t exist.
Narrow job. An agent that “helps with everything” helps with nothing.
Output format specified. The skill consumes a known format; the agent produces it.
Refusal patterns named. What the agent won’t do is as important as what it will.