Skip to content

Agents

Worclaude installs Claude Code agent definitions as Markdown files in .claude/agents/. Each agent has a specific model, isolation mode, and purpose. Universal agents are always installed. Optional agents are selected during worclaude init based on project type. See Claude Code Integration for details on how agents register with the runtime.

Agent Properties

Every agent file includes a YAML frontmatter block. The name and description fields are required -- without description, agents are invisible to Claude Code and will not appear in /agents or be available for routing.

PropertyRequiredValuesPurpose
nameYesstringAgent identifier used for routing
descriptionYesstringWhat the agent does -- required for visibility
modelYesopus, sonnet, haiku, inheritWhich Claude model the agent uses
isolationNonone, worktreeWhether the agent works in a git worktree
disallowedToolsNostring[]Tools the agent cannot use (enforced by runtime)
toolsNostring[]Tool allowlist -- alternative to disallowedTools
backgroundNobooleanRun asynchronously without blocking the user
maxTurnsNonumberMaximum conversation turns before auto-stop
omitClaudeMdNobooleanSkip loading CLAUDE.md for this agent
memoryNoprojectMemory scope for cross-session learning
effortNolow, medium, high, or integerControls token spend for the agent
colorNostringAgent color in the UI (e.g., orange, red)
permissionModeNoacceptEdits, bypassPermissions, default, dontAsk, planPer-agent permission mode override
mcpServersNostring[] or object[]MCP servers this agent requires
hooksNoobjectPer-agent hooks that register when the agent starts

tools vs disallowedTools

Use tools (allowlist) when you know exactly which tools an agent needs: tools: ['Read', 'Bash', 'Glob', 'Grep']. Use disallowedTools (denylist) when the agent needs most tools but should be blocked from a few: disallowedTools: ['Edit', 'Write']. Both are enforced by the runtime — the agent physically cannot access restricted tools.

WARNING

Without a description field, agents are invisible to Claude Code. They will not appear in /agents and cannot be routed to. Run worclaude doctor to detect agents with missing descriptions.

Model Selection

Each agent is assigned a model based on the complexity and nature of its task:

ModelToken CostBest ForAgents Using It
OpusHighestDeep judgment, architectural decisions, security analysisplan-reviewer, api-designer, auth-auditor, security-reviewer, prompt-engineer
SonnetMediumImplementation, code changes, testing, documentationcode-simplifier, test-writer, verify-app, database-analyst, ui-reviewer, ci-fixer, docker-helper, deploy-validator, performance-auditor, bug-fixer, refactorer, build-fixer, e2e-runner, doc-writer, data-pipeline-reviewer, ml-experiment-tracker
HaikuLowestNarrow validation, formatting, simple checksbuild-validator, style-enforcer, dependency-manager, changelog-generator

Isolation Modes

ModeHow It WorksWhen Used
noneAgent works in the current working directory alongside the user. Changes are made directly to the working tree.Read-only review tasks, validation, analysis. No risk of conflicting with user changes.
worktreeAgent creates an isolated git worktree from the current branch. All changes happen in the worktree. User reviews and merges results afterward.Code modification tasks (writing tests, fixing bugs, refactoring). Prevents conflicts with uncommitted user work. Requires clean git state to create the worktree.

Agents with worktree isolation: code-simplifier, test-writer, verify-app, ci-fixer, bug-fixer, refactorer, build-fixer, e2e-runner, doc-writer.


Universal Agents

These 5 agents are installed with every Worclaude project. They cover the core development workflow: planning, coding, testing, building, and verifying.

plan-reviewer

Modelopus
Isolationnone
Invoked by/review-plan slash command
disallowedToolsEdit, Write, NotebookEdit, Agent
omitClaudeMdtrue

Acts as a senior staff engineer reviewing an implementation plan. Challenges assumptions, identifies ambiguity, checks for missing verification steps, and ensures the plan is specific enough for one-shot implementation. Validates alignment with SPEC.md. Does not approve plans that are vague or lack verification. Read-only: cannot modify files.

code-simplifier

Modelsonnet
Isolationworktree

Reviews recently changed code and improves it. Finds and eliminates duplication, identifies reuse opportunities with existing code, simplifies complex logic, and ensures consistency with project patterns. Makes changes directly in a worktree and runs tests after each change. Commits improvements separately from feature work.

test-writer

Modelsonnet
Isolationworktree
memoryproject

Writes comprehensive tests for recently changed code. Covers unit tests for individual functions, integration tests for component interactions, edge cases (null, empty, boundary values), and error paths. Follows the project's testing patterns from .claude/skills/testing/SKILL.md. Aims for meaningful coverage rather than 100% line coverage. Has project memory for learning test patterns across sessions.

build-validator

Modelhaiku
Isolationnone
backgroundtrue

Validates that the project builds and passes all checks: build command, full test suite, linter, and type checker (if applicable). Reports failures with clear error messages. Does not fix issues -- reports them so the main session can address them. Runs in the background so you can continue working while it validates.

verify-app

Modelsonnet
Isolationworktree
Invoked by/verify slash command (indirectly)
backgroundtrue

Tests the actual running application behavior, not just unit tests. Starts the application, tests changed functionality end-to-end, verifies behavior matches the specification, checks for regressions in related features, and tests error handling and edge cases in the running app. Reports specific pass/fail for each verification step. Runs in the background.


Optional Agents

20 optional agents organized into 6 categories. During worclaude init, categories are recommended based on the selected project type.

Backend

AgentModelIsolationPurpose
api-designeropusnoneReviews API design for RESTful conventions, consistency, and completeness.
database-analystsonnetnoneReviews database schemas, queries, migrations, and indexing strategies.
auth-auditoropusnoneAudits authentication and authorization implementations for security flaws.

Frontend

AgentModelIsolationPurpose
ui-reviewersonnetnoneReviews UI for visual consistency, accessibility, and responsive behavior.
style-enforcerhaikunoneEnsures design system compliance across components and styles.

DevOps

AgentModelIsolationPurpose
ci-fixersonnetworktreeDiagnoses and fixes CI/CD pipeline failures.
docker-helpersonnetnoneReviews Docker configurations for best practices and optimization.
deploy-validatorsonnetnoneValidates deployment readiness including configs, environment variables, and dependencies.
dependency-managerhaikunoneReviews dependency health, outdated packages, and security advisories.

Quality

AgentModelIsolationPurpose
bug-fixersonnetworktreeDiagnoses root causes and implements fixes for reported bugs.
security-revieweropusnoneReviews code for security vulnerabilities, injection risks, and unsafe patterns.
performance-auditorsonnetnoneAnalyzes code for performance bottlenecks, memory leaks, and inefficiencies.
refactorersonnetworktreeRefactors code to improve maintainability, readability, and structure.
build-fixersonnetworktreeDiagnoses and fixes build failures, test failures, lint errors, and type errors.
e2e-runnersonnetworktreeWrites and runs end-to-end tests for critical user journeys.

Documentation

AgentModelIsolationPurpose
doc-writersonnetworktreeWrites and updates project documentation including READMEs, guides, and API docs.
changelog-generatorhaikunoneGenerates changelog entries from commit history.

Data / AI

AgentModelIsolationPurpose
data-pipeline-reviewersonnetnoneReviews data pipeline correctness, schema consistency, and transformation logic.
ml-experiment-trackersonnetnoneReviews ML experiment reproducibility, hyperparameter tracking, and result logging.
prompt-engineeropusnoneReviews and improves LLM prompts for clarity, specificity, and effectiveness.

Category Recommendations by Project Type

During worclaude init, agent categories are pre-selected based on the chosen project type. Users can override these selections.

Project TypeRecommended Categories
Full-stack web applicationBackend, Frontend, Quality, Documentation
Backend / APIBackend, Quality
Frontend / UIFrontend, Quality
CLI toolQuality, Documentation
Data / ML / AIData / AI, Backend
Library / PackageQuality, Documentation
DevOps / InfrastructureDevOps

The following table shows the specific agents recommended for each project type:

Project TypeRecommended Agents
Full-stack web applicationui-reviewer, api-designer, database-analyst, security-reviewer, bug-fixer, doc-writer, e2e-runner
Backend / APIapi-designer, database-analyst, security-reviewer, auth-auditor, bug-fixer, performance-auditor, build-fixer
Frontend / UIui-reviewer, style-enforcer, performance-auditor, bug-fixer, e2e-runner
CLI toolbug-fixer, doc-writer, dependency-manager, build-fixer
Data / ML / AIdata-pipeline-reviewer, ml-experiment-tracker, prompt-engineer, database-analyst
Library / Packagedoc-writer, dependency-manager, performance-auditor, refactorer, changelog-generator
DevOps / Infrastructureci-fixer, docker-helper, deploy-validator, dependency-manager

Summary Table (All 25 Agents)

AgentCategoryModelIsolationDescription
plan-reviewerUniversalopusnoneReviews implementation plans as a staff engineer
code-simplifierUniversalsonnetworktreeEliminates duplication, simplifies logic
test-writerUniversalsonnetworktreeWrites comprehensive tests for changed code
build-validatorUniversalhaikunoneValidates build, tests, lint without fixing
verify-appUniversalsonnetworktreeEnd-to-end application testing
api-designerBackendopusnoneReviews API design for RESTful conventions
database-analystBackendsonnetnoneReviews database schemas and queries
auth-auditorBackendopusnoneAudits authentication and authorization
ui-reviewerFrontendsonnetnoneReviews UI for consistency and accessibility
style-enforcerFrontendhaikunoneEnsures design system compliance
ci-fixerDevOpssonnetworktreeDiagnoses and fixes CI/CD failures
docker-helperDevOpssonnetnoneReviews Docker configs for best practices
deploy-validatorDevOpssonnetnoneValidates deployment readiness
dependency-managerDevOpshaikunoneReviews dependency health and updates
bug-fixerQualitysonnetworktreeDiagnoses and fixes bugs
security-reviewerQualityopusnoneReviews code for security vulnerabilities
performance-auditorQualitysonnetnoneAnalyzes code for performance issues
refactorerQualitysonnetworktreeRefactors code to improve maintainability
build-fixerQualitysonnetworktreeDiagnoses and fixes build failures
e2e-runnerQualitysonnetworktreeWrites and runs end-to-end tests
doc-writerDocumentationsonnetworktreeWrites and updates documentation
changelog-generatorDocumentationhaikunoneGenerates changelog from commits
data-pipeline-reviewerData / AIsonnetnoneReviews data pipeline correctness
ml-experiment-trackerData / AIsonnetnoneReviews ML experiment reproducibility
prompt-engineerData / AIopusnoneReviews and improves LLM prompts

Runtime Properties

The following table shows which agents have special runtime properties beyond the standard model/isolation fields:

AgentRead-OnlyBackgroundMemoryomitClaudeMd
plan-reviewerYesYes
test-writerproject
build-validatorYes
verify-appYes
api-designerYes
database-analystYes
ui-reviewerYes
security-reviewerprojectYes
performance-auditorYes
dependency-managerYes
deploy-validatorYes
e2e-runnerYes
doc-writerproject
changelog-generatorYesYes
data-pipeline-reviewerYes
ml-experiment-trackerYes

Read-Only = has disallowedTools preventing file modifications. Background = runs asynchronously via background: true. Memory = has cross-session learning via memory field. omitClaudeMd = does not load CLAUDE.md (narrower focus).

Agents not listed have no special runtime properties beyond model and isolation.


Agent File Location

All agent files are installed to .claude/agents/ as flat Markdown files:

.claude/agents/
  plan-reviewer.md          # universal
  code-simplifier.md        # universal
  test-writer.md            # universal
  build-validator.md        # universal
  verify-app.md             # universal
  api-designer.md           # optional (if selected)
  bug-fixer.md              # optional (if selected)
  ...

Agent files can be customized after installation. The worclaude diff command tracks modifications, and worclaude upgrade preserves customizations by saving new versions as .workflow-ref.md files.


See Also

Built with best practices from Boris Cherny's Claude Code tips.