[Feature]: Repository-Defined Custom Agents Support

[theonetheycallneo]

Feature Area

UI/UX (User Interface)

Priority

Critical for my use case

Problem Statement

AutoMaker's Model Defaults allow assigning different AI models to task types (Quick Tasks, Validation Tasks, Generation Tasks), but modern agentic workflows require more than model selection—they need specialized agents with distinct personas, system prompts, and tool configurations.

Currently there's no way to:

• Define project-specific agents with custom system prompts and skills
• Version-control agent configurations alongside code
• Assign tasks to specialized agents (planner, dev, qa, ux) based on task type
• Share consistent agent definitions across team members

This limits AutoMaker's effectiveness for teams adopting multi-agent methodologies where different phases of work benefit from purpose-built agent personas.

Proposed Solution

Add support for discovering and utilizing custom agent definitions from .claude/agents/, .github/agents/, or .automaker/agents/ directories within repositories.

Agent Definition Format

# .claude/agents/qa-engineer.yaml
name: QA Engineer
description: Specializes in testing, validation, and quality assurance
model: claude-sonnet  # optional model override

system_prompt: |
  You are a senior QA engineer focused on comprehensive testing.
  Always consider edge cases, error handling, and security implications.
  Generate test cases before approving any implementation.

skills:
  - testing
  - code-review
  - security-audit

tools:
  - file_read
  - file_write
  - terminal
  - browser

task_types:
  - validation
  - code-review
  - test-generation

UI Integration

Extend the Model Defaults panel to show an agent dropdown alongside model selection:

┌─────────────────────────────────────────────────────────────┐
│ Task Type          │ Agent              │ Model (fallback)  │
├─────────────────────────────────────────────────────────────┤
│ Code Generation    │ [Dev Agent ▼]      │ Claude Sonnet     │
│ Code Review        │ [QA Agent ▼]       │ Claude Sonnet     │
│ Spec Writing       │ [Planner Agent ▼]  │ Claude Opus       │
│ Commit Messages    │ [None ▼]           │ Claude Haiku      │
└─────────────────────────────────────────────────────────────┘

Alternatives Considered

1. Global agent definitions only (in Global Settings)

• Pros: Simpler implementation, single source of truth
• Cons: Not version-controlled, can't vary per project, no team sharing

2. Prompt Customization per task type (existing feature)

• Pros: Already available
• Cons: Doesn't support full agent personas with skills/tools, no reusability across task types

3. External agent management tool

• Pros: Separation of concerns
• Cons: Context switching, sync issues, doesn't leverage AutoMaker's existing task/model infrastructure

4. MCP Server-only approach

• Pros: Powerful tool integration
• Cons: Overkill for simple persona/prompt customization, steeper learning curve

Use Cases

1. BMAD-style multi-agent workflow
A team using the BMAD-METHOD defines agents for each phase:

• Planner agent creates specs and breaks down features
• Dev agent implements code
• QA agent validates and writes tests
• Tasks flow through agents based on type

2. Specialized code review
A security-focused project defines a security-reviewer agent with prompts emphasizing vulnerability detection, OWASP guidelines, and dependency auditing. All code review tasks automatically use this agent.

3. Consistent team standards
A distributed team commits their agent definitions to the repo. New contributors get the same agent behaviors without manual setup, ensuring consistent code style, documentation quality, and testing approaches.

4. Project-specific context
A fintech project defines agents pre-loaded with domain knowledge about compliance requirements, PCI-DSS, and financial calculations. A gaming project has agents tuned for performance optimization and graphics pipeline patterns.

5. Incremental adoption
Teams start with one custom agent for their biggest pain point (e.g., test generation), then expand to other agents over time. Existing model defaults continue working for undefined task types.

Mockups/Screenshots

Mockup concept:

Example Directory Structure:

Additional Context

Related Projects

• BMAD-METHOD - Multi-agent workflow methodology
• Cursor Rules - .cursorrules for project-specific AI behavior
• Aider Conventions - .aider.conf.yml for repo-specific settings
• Claude Projects - Custom instructions per project

Example Directory Structure

project/
├── .claude/
│   └── agents/
│       ├── planner.yaml
│       ├── developer.yaml
│       ├── qa-engineer.yaml
│       └── ux-designer.yaml
└── src/

Checklist

• I have searched existing issues to ensure this feature hasn't been requested already
• I have provided a clear description of the problem and proposed solution