Skip to content

Agent Creator

Enable human-AI collaboration to design and create specialized Agents through iterative refinement.

Core Capabilities

Agent Creator provides guided Agent design through conversation:

Capability Benefit Example
Requirements Gathering Agent asks clarifying questions to understand needs User says "create coding agent" → Agent asks about languages, tools, use cases
Specification Guidance Provides complete design specification with parameters explained Learn what fields like instruction, max_actions, mcp mean and how to configure them
Configuration as Files Agent designs are JSON files users can review and edit agent/code_reviewer.json created in conversation, visible in file system
Export Existing Agents Study successful Agent configurations Export data_analyst config to learn patterns for new analytics Agent
Iterative Refinement Modify designs based on feedback until satisfied Draft → User feedback → Adjustment → Repeat until perfect

Collaboration Pattern:

  1. User states need in natural language
  2. Agent Creator asks clarifying questions (languages, tools, scenarios)
  3. Gets specification to understand all configuration options
  4. Designs Agent config based on user requirements
  5. Creates JSON file for user to review
  6. User provides feedback, Agent adjusts design
  7. Repeat until user approves
  8. Registers finalized Agent to system

Available Operations

Operation Purpose When to Use
get_agent_spec Get complete Agent design specification Starting a new Agent design to understand parameters
export_agent_to_document Export existing Agent config to JSON file Learning from successful Agents or creating variations
register_agent Save Agent design to system After user approves final configuration

Note: Agent Creator works with File Tool to create/edit JSON configuration files.

Common Workflows

Creating New Agent

Stage Action Details
1. Requirements Agent asks questions "What tasks will this Agent perform?", "What tools does it need?", "Which model?"
2. Specification Get design guide get_agent_spec retrieves parameter explanations and examples
3. Design Draft configuration Agent designs card, instruction, selects model, tools, category
4. Review Create file & discuss create_file("agent/reviewer.json", ...) → User reviews design
5. Refine Iterate on feedback User: "Add Python focus" → Agent updates with update_file
6. Register Save to system User approves → register_agent("agent/reviewer.json")

Optimizing Existing Agent

Stage Action Details
1. Export Get current config export_agent_to_document("code_analyzer") → Creates agent/code_analyzer.json
2. Review User identifies changes "Make it more detailed", "Add security focus"
3. Modify Update configuration Agent edits JSON using update_file or rewrite_file
4. Refine Iterate until satisfied Multiple rounds of user feedback and adjustments
5. Update Re-register register_agent("agent/code_analyzer.json") → Updates existing Agent

Creating Agent Variations

Stage Action Result
1. Export base export_agent_to_document("python_expert") agent/python_expert.json
2. Duplicate create_file("agent/javascript_expert.json") Copy of Python expert config
3. Customize Modify language-specific details Change instruction, tools, tags for JavaScript
4. Register register_agent("agent/javascript_expert.json") New JavaScript expert Agent

Design Principles

The specification guides Agents to follow these principles:

Clear Capabilities: Agent should have well-defined purpose and scope

Detailed Descriptions: card field explains what Agent does and when to use it

Precise Instructions: instruction accurately defines Agent's role and behavior

Appropriate Tools: Select tools matching Agent's domain (code tools for reviewers, file tools for analysts)

Semantic Naming: Clear names and tags help discovery

Reasonable Limits: Set max_actions based on task complexity

Integration with Other Tools

+ File Tool: Agent configs are JSON files managed with file operations - create_file to create initial design - read_file to review current config - update_file for targeted changes - rewrite_file for major revisions

+ Sub-agent: Created Agents can be used via Sub-agent tool - Design specialist Agents - Delegate tasks to them - Iterate on their configurations based on performance