agent-builder

# agent-builder Give it a task and a framework. It generates a complete, working agent project with the right structure, dependencies, and best practices. ## Inputs | input | default | notes | |---|---|---| | task | required | What the agent should do (e.g., "research assistant that searches the web") | | framework | ask user | `autogen`, `crewai`, `pydanticai`, `smolagents`, or `openai` | | project_name | derived | Snake_case name derived from task | | model | gpt-5.4-mini | Default LLM to use | | include_tools | true | Generate custom tools if the task needs them | ## Framework Selection Guide When the user doesn't specify a framework, recommend based on their task: | Task Type | Recommended Framework | Why | |-----------|----------------------|-----| | Multi-agent collaboration | AutoGen | Best for complex agent-to-agent workflows | | Role-based teams | CrewAI | Built for crew metaphor with roles/tasks | | Type-safe, structured output | PydanticAI | Pydantic integration, strong typing | | Lightweight, code execution | Smolagents | Minimal overhead, code agents | | Simple single agent | OpenAI Agents | Straightforward, OpenAI native | ## Process ### 1. Gather requirements Ask the user: 1. "What task should the agent perform?" (required) 2. "Which framework? (autogen/crewai/pydanticai/smolagents/openai)" - if not specified, recommend 3. "Any specific requirements? (tools, APIs, output format)" ### 2. Validate the framework choice Check `reference/frameworks.md` for the framework's current API patterns. Each framework has different conventions - follow them exactly. ### 3. Design the agent architecture Based on the task: - Single agent: Simple tasks, one responsibility - Multi-agent: Complex tasks needing collaboration - Crew/Team: Tasks with distinct roles (research, write, review) Identify required tools: - Web search: DuckDuckGo, Tavily, Serper - Code execution: Docker sandbox, local executor - File operations: Read/write files - API calls: Custom integrations - Database: Vector stores, SQL ### 4. Generate the project Use the appropriate template from `templates/`. Generate: 1. **README.md** - How to set up and run 2. **requirements.txt** - All dependencies with versions 3. **main.py** - Entry point 4. **Agent files** - Framework-specific agent definitions 5. **Tools** - Custom tools if needed 6. **Config** - Any configuration files ### 5. Validate the code Before presenting to user: - Check imports are correct for the framework version - Ensure all tools are properly defined - Verify the entry point runs without syntax errors ## Templates Each template in `templates/` contains: - Current import patterns for that framework - Project structure conventions - Example agent and tool definitions - Common gotchas to avoid ## Rules - **Use current API patterns.** AutoGen v0.7+ uses async/await and `autogen_agentchat`. CrewAI uses decorators and YAML config. Check `reference/frameworks.md` for current patterns. - **Pin dependency versions.** Always include version pins in requirements.txt. - **Generate working code.** Every file should be syntactically correct and runnable. - **Follow framework conventions.** CrewAI uses YAML for agents/tasks, AutoGen uses Python classes, PydanticAI uses decorators. Match the framework's idioms. - **Include error handling.** Wrap API calls in try/except, handle missing API keys gracefully. - **Document everything.** README should explain setup, configuration, and usage. - **No placeholders.** Generate complete, working code. If something needs user input, use environment variables with clear documentation.