Agent Creator

Agent Creator

The Agent Creator skill is the counterpart to skill-creator, specifically designed for the rapid generation and iterative improvement of Cursor agents. It ensures every agent complies with the canonical schemas/agent.schema.json and follows the factory core patterns.

Standard Structure

All agents follow this standardized markdown format in {directories.agents}/*.md:

1. YAML Frontmatter: Mandatory fields (name, description, type, version, domain, skills, knowledge, tools, workflows, blueprints).
2. Purpose: High-level goal.
3. Philosophy: Guiding quote or principle.
4. Triggers: When the agent should activate.
5. Workflows: Mermaid diagram of the agent's logic.
6. Rules: Core constraints.

Iterative Improvement Loop

Like skill-creator, this skill supports an evaluation loop:

1. Generate: Draft an agent based on a pattern or requirement.
2. Evaluate: Run the agent against test queries (using scripts/run_eval.py).
3. Analyze: Use evaluation agents (agents/analyzer.md, etc.) to identify weaknesses.
4. Improve: Refine the agent's description and rules (using scripts/run_loop.py).

Automation Scripts

Located in scripts/:

• run_loop.py: Orchestrates the improvement cycle.
• quick_validate.py: Validates agent frontmatter against agent.schema.json.
• package_skill.py: (Internal use) Packages agent definitions.

Usage

Use this skill to refactor existing agents or create new specialized agents for the Factory.

When to Use

• When creating a new agent definition from scratch
• When updating an existing agent to comply with the Factory schema
• When a user wants to design a domain-specialist agent for a specific use case
• When another agent or workflow requests a properly structured agent definition

Prerequisites

• Access to Factory schemas in schemas/
• Understanding of available skills and topologies (chain, parallel, routing, etc.)
• The quick_validate.py script must be available for schema validation

Process

1. TDD Phase (RED):
   • Write a failing verification test in tests/agents/.
   • Verify RED: Run the test and confirm failure.
2. Implementation Phase (GREEN):
   • Identify relevant skills and knowledge.
   • Draft the agent .md file with minimal content to pass the test.
3. Validation Phase:
   • Validate against the agent schema using quick_validate.py.
   • Verify GREEN: Run the verification test again to confirm it passes.
4. REFACTOR & Sync:
   • Iterate with user feedback.
   • Mandatory Sync: Run conda run -p D:\Anaconda\envs\cursor-factory python scripts/sync_global_workflows.py.

Best Practices

• Always define a clear domain: and ype: agent in frontmatter
• Only reference skills that actually exist in the Factory
• Keep agent scope focused — prefer depth over breadth
• Validate the skills: list against the skill catalog
• Document axioms and decision gates for clarity