custom-agent-definitions

# Custom Agent Definitions Expert knowledge for defining and configuring custom agents in Claude Code. For full worked YAML examples (isolated research agent, read-only explorer, complete security auditor, plugin layout, common patterns), see [REFERENCE.md](REFERENCE.md). ## When to Use This Skill | Use this skill when... | Use agent-teams instead when... | |---|---| | Authoring a new `.md` agent definition file in `.claude/agents/` | Spawning multiple already-defined agents that coordinate as a team | | Configuring a single agent's `model`, `allowed-tools`, or `isolation` | Setting up a lead/teammate architecture with a shared task list | | Constraining tool access for a specialised read-only or write-restricted agent | Sequencing parallel work across worktrees (see parallel-agent-dispatch) | | Writing the system prompt that defines what one agent does | Auditing existing agent definitions for security (see meta-audit) | ## Core Concepts **Custom agents** let you define specialized agent types beyond the built-in ones (Explore, Plan, Bash, etc.). Each can have its own model, tools, and isolation settings. They are defined in `.claude/agents/` or via plugin `agents/` directories, with YAML frontmatter + a markdown system prompt: ```yaml --- name: my-custom-agent description: What this agent does model: sonnet allowed-tools: Bash, Read, Grep, Glob --- # Agent System Prompt Instructions and context for the agent... ``` ## Key Fields ### Context: isolated by default A named agent always starts in a fresh context: its own system prompt, the brief the caller writes, CLAUDE.md, and any preloaded `skills:`. It does not see the caller's conversation history, and no agent frontmatter field changes that. | Want | Use | |------|-----| | A delegate that keeps verbose work out of the main context | A named agent — isolation is the default, no field needed | | A subagent that already knows the conversation so far | `subagent_type: "fork"` on the `Agent` call (fork mode is off under `-p` unless `CLAUDE_CODE_FORK_SUBAGENT=1`) | `context: fork` is a **skill** frontmatter field that runs a skill body in a new subagent. On an agent it is not a documented field, and Claude Code ignores it without an error. See [REFERENCE.md → Isolated research agent](REFERENCE.md#isolated-research-agent) and `.claude/rules/agent-development.md` § Context Isolation. ### Tool Access (allowed vs disallowed) | Field | Purpose | Behavior | |-------|---------|----------| | `allowed-tools` | Whitelist of permitted tools | Agent can ONLY use these tools | | `disallowedTools` | Blacklist of forbidden tools | Agent can use all tools EXCEPT these | Use `disallowedTools` for read-only agents, restricting dangerous capabilities, and sandboxing. The two combine — an explicit whitelist plus a safety blacklist. See [REFERENCE.md → Read-only explorer](REFERENCE.md#read-only-explorer-disallowedtools). ### Agent Field for Delegation The `agent` field specifies which agent type to use when delegating via the Agent tool, letting commands and skills name a preferred agent type: ```yaml agent: security-auditor ``` ## Agent Configuration Fields Reference | Field | Type | Description | |-------|------|-------------| | `name` | string | Agent identifier | | `description` | string | What the agent does | | `model` | string | `opus`, `sonnet`, `haiku`, `fable`, `inherit`, or a full model ID | | `effort` | string | `low`, `medium`, `high`, `xhigh`, `max` — overrides the session effort while this agent runs; default inherits. The cost lever for mechanical delegates | | `permissionMode` | string | `default`, `acceptEdits`, `dontAsk`, `bypassPermissions`, or `plan` | | `maxTurns` | number | Maximum agentic turns before agent stops | | `background` | bool | Set `true` to always run as a background task | | `memory` | string | Persistent memory scope: `user`, `project`, or `local` | | `skills` | list | Skill names to preload into agent context at startup | | `mcpServers` | list | MCP server names available to this agent | | `tools` | list | Tools the agent can use (in agents/ dir; use `allowed-tools` in skills) | | `disallowedTools` | list | Tools the agent cannot use | | `created` / `modified` / `reviewed` | date | Lifecycle dates | ## Best Practices 1. **Principle of least privilege** — grant only the tools the agent needs. 2. **Rely on default isolation** — a named agent never sees the caller's conversation, so exploratory work stays out of the main context without any field. Only documented agent fields take effect; an unrecognized key, such as a skill's `context:`, is ignored without an error. 3. **Combine allowed + disallowed** — explicit whitelist with a safety blacklist. 4. **Clear descriptions** — describe what the agent does and its boundaries. 5. **Model and effort** — `model: opus` is the floor for any agent whose output re-enters the main loop (a weaker delegate degrades everything downstream; `scripts/check-agent-model.sh` enforces it for plugin agents). `fable` is sanctioned for the hardest delegated reasoning. Tune cost with `effort:` (`low` for mechanical work), not by downgrading the model. The one exception is the `cold-read-gate` haiku reader, which is a measurement instrument, not a delegate. See `.claude/rules/agent-development.md` § "Model Selection for Agents" (repo) and `~/.claude/rules/agent-and-tool-selection.md` (user-global). 6. **Report failures loudly** — a dispatched agent that hits a wall must say so in its final message, never a one-word summary like `Terminal.` / `Done.` / `Stopped.` On a blocker it should commit and push its in-progress work, open a draft PR, and state exactly what stopped it and which tools were denied. A one-word surrender is indistinguishable from success to the orchestrator, so the work is silently cleaned up and lost (issue [#1422](https://github.com/laurigates/claude-plugins/issues/1422)). See `parallel-agent-dispatch` → "Loud-failure contract" for the dispatch-prompt form every brief should carry. 7. **Prefer a Skill-less agentType for read-only fan-out** — an agent that only reads files and emits structured output should NOT carry the `Skill` tool. Every `Skill`-bearing agent pays a ~25k-token `skill_listing` + `deferred_tools_delta` context tax before its first tool call, which can push read-heavy fan-out subagents over their context window. Use a lean read-only agent (e.g. `agents-plugin:review`) instead. See `parallel-agent-dispatch` → "Skill-less agentType for Read-Only Fan-Out" (issues [#1549](https://github.com/laurigates/claude-plugins/issues/1549) / [#1550](https://github.com/laurigates/claude-plugins/issues/1550)). Worked YAML for each practice is in [REFERENCE.md → Best-practice snippets](REFERENCE.md#best-practice-snippets). ## Quick Reference ### Context Inheritance | Dispatch | Sees the caller's conversation | Use Case | |----------|--------------------------------|----------| | Named agent (`subagent_type: "<name>"`) | No — brief only | Research, review, tool-bounded work | | `subagent_type: "fork"` | Yes — the whole conversation | Side task that needs the prior context | ### Tool Restriction Patterns | Pattern | Fields | |---------|--------| | Whitelist only | `allowed-tools: Tool1, Tool2` | | Blacklist only | `disallowedTools: Tool1, Tool2` | | Combined | Both fields specified | ## Related - [REFERENCE.md](REFERENCE.md) — full worked YAML examples and snippets - `agent-teams` — multi-agent coordination via the implicit team - `parallel-agent-dispatch` — worktree preflight, scope budgets, loud-failure contract - `meta-audit` — auditing existing agent definitions for security/completeness - `.claude/rules/agent-development.md` — agent lifecycle and field semantics