Agentic Development Skill For building autonomous AI agents that perform multi step tasks with tools. Sources: [Claude Agent SDK](https://docs.anthropic.com/en/docs/agents and tools/claude agent sdk) [Anthropic Claude Code Best Practices](https://www.anthropic.com/engineering/claude code best practices) [Pydantic AI](https://ai.pydantic.dev/) [Google Gemini Agent Development](https://developers.googleblog.com/en/building agents google gemini open source frameworks/) [OpenAI Building Agents](https://developers.openai.com/tracks/building agents/) Framework Selection by Language Language/Framework Default Why Python Pydantic AI Type safe, Pydantic validation, multi model, production ready Node.js / Next.js Claude Agent SDK Official Anthropic SDK, tools, multi agent, native streaming Python: Pydantic AI (Default) Node.js / Next.js: Claude Agent SDK (Default) Core Principle Plan first, act incrementally, verify always. Agents that research and plan before executing consistently outperform those that jump straight to action. Break complex tasks into verifiable steps, use tools judiciously, and maintain clear state throughout execution. Agent Architecture Three Components (OpenAI) Project Structure Workflow Pattern: Explore Plan Execute Verify 1. Explore Phase 2. Plan Phase (Critical) 3. Execute Phase 4. Verify Phase Tool Design Tool Definition Pattern Tool Implementation Prefer Built in Tools (OpenAI) Multi Agent Patterns Single Agent (Default) Use one agent for most tasks. Multiple agents add complexity. Agent as Tool Pattern (OpenAI) Handoff Pattern (OpenAI) When to Use Multiple Agents Separate task domains with non overlapping tools Different authorization levels needed Complex workflows with clear handoff points Parallel execution of independent subtasks Memory & State Conversation Memory Persistent Memory Guardrails & Safety Multi Layer Protection (OpenAI) Scope Enforcement (OpenAI) Model Selection Match Model to Task Task Complexity Recommended Model Notes Simple, fast gpt 5 mini, claude haiku Low latency General purpose gpt 4.1, claude sonnet Balance Complex reasoning o4 mini, claude opus Higher accuracy Deep planning gpt 5 + reasoning, ultrathink Maximum capability Gemini Specific Claude Specific (Thinking Modes) Testing Agents Unit Tests (Tools) Behavior Tests (Agent Decisions) Evaluation Tests Pydantic AI Patterns (Python Default) Project Structure (Python) Agent with Tools Structured Output with Validation python\n{code}\n Multi Agent Coordination Streaming Responses Testing Agents Skills Pattern (Anthropic) Skill Structure instructions.md Example Loading Skills Dynamically Anti Patterns No planning before execution Agents that jump to action make more errors Monolithic agents One agent with 50 tools becomes confused No verification Agents must verify their own work Hardcoded tool sequences Let the model decide tool order Missing guardrails All agents need safety boundaries No state management Lose context across tool calls Testing only happy paths Test failures and edge cases Ignoring model differences Reasoning models need different prompts No cost tracking Agentic workflows can be expensive Full automation without oversight Human in the loop for critical actions Quick Reference Agent Development Checklist [ ] Define clear agent scope and boundaries [ ] Design tools with explicit schemas and risk levels [ ] Implement explore plan execute verify workflow [ ] Add multi layer guardrails [ ] Set up conversation and persistent memory [ ] Write behavior and evaluation tests [ ] Configure appropriate model for task complexity [ ] Add human in the loop for high risk operations [ ] Monitor token usage and costs [ ] Document skills and instructions Thinking Triggers (Claude) Gemini Settings

agentic-development