Agents

This documentation was generated with the assistance of AI. Please report any inaccuracies.

An agent is a sub-instance of Claude Code, run via the Agent tool, with its own separate context window — spawned either to search the codebase without derailing the current conversation, or to isolate a sub-task (a verbose report, an entire multi-step skill run, a batch of independent summarization work) so only a distilled result lands back in the calling skill’s own context.

This repository defines three custom agents under .claude/agents/, each with its own detail page — Gate Runner, Isolated Skill Executor, and Change Summarizer — for shapes that recurred often enough across this catalog’s skills to be worth centralizing. This page instead documents the two built-in agent types this catalog still relies on for everything that doesn’t fit one of those three shapes.

Explore

A fast, read-only search agent for locating code — finding files by pattern, grepping for a symbol, or answering "where is X defined". It cannot edit files.

Used by

  • Explore — Step 3 uses it to build a picture of the codebase’s structure (module layout, core abstractions, control flow) before reporting findings, in place of manually reading every file.

  • Setup Antora — Step 7 uses it to locate the top-level source layout and main public entry points before writing `index.adoc’s repository overview, instead of reading the whole source tree directly.

  • Setup Readme — Step 7 uses it the same way, to locate the main source entry points before writing the README’s "how it works" example.

Why this one specifically

Each of these is read-only research over a codebase whose size and shape aren’t known in advance — iru-explore’s whole job, and a narrower, bounded version of the same need in `iru-setup-antora/iru-setup-readme’s own Step 7. The `Explore agent is purpose-built for exactly that: broad, targeted search that stays out of the main conversation’s context window, without the ability to accidentally make an edit during what’s meant to be a non-destructive pass.

general-purpose

Claude Code’s default agent type — used whenever a skill spawns a sub-agent via the Agent tool without naming a specific type. It has the same tool access as the main conversation.

Used by

  • Generate Skill Docs — Step 3 spawns one per skill and per documented agent found, each reading only its own file and returning a compact structured summary (purpose, inputs, outputs, steps, invokes/spawns edges) instead of this skill’s own conversation reading every SKILL.md in the catalog directly; separately, it also spawns one (as does Update Docs) only when no Antora documentation module exists yet, to run the iru-setup-antora skill (briefed with the context of what’s about to follow) before continuing with the rest of their own work.

  • PR Description — Step 2 spawns one, given the branch/base-branch names, to read the full diff and recent commit messages and draft the PR description, so the raw diff stays out of this skill’s own context and only the drafted text comes back.

Why this one specifically

Bootstrapping a documentation module mid-run, extracting one skill’s metadata, and drafting a description from a diff are each one-off, occasional delegations — not a shape common enough across the catalog to justify a dedicated custom agent definition the way Gate Runner, Isolated Skill Executor, and Change Summarizer were. The plain default is the right choice here: nothing about the delegation needs restricted tool access or a specialized system prompt.

Adding a custom agent to this catalog

If a future need doesn’t fit either built-in type above or one of this catalog’s three custom agents, define it under .claude/agents/<name>.md following the same frontmatter convention Claude Code uses for agent definitions, then re-run the iru-generate-skill-docs skill so this page and AI Catalog's dependency graph pick it up.