Agentic AI Dev

Iron Law

NO AGENT GRAPH WITHOUT AN ITERATION LIMIT AND A HUMAN-IN-THE-LOOP CHECKPOINT — unbounded loops and silent runaway agents are production incidents

Agentic AI Development Skill — Python 3.14 + LangChain + LangGraph + FastAPI

Quick Scaffold

uv init my-agent-service && cd my-agent-service
uv add "langchain-core>=1.2.8" "langchain-anthropic>=1.3.0" "langchain-openai>=1.1.0" "langgraph>=1.0.7" \
  "fastapi>=0.135.2" "uvicorn[standard]" pydantic pydantic-settings \
  langsmith prometheus-client structlog httpx asyncpg \
  "langgraph-checkpoint-postgres>=3.0.0"
uv add --dev pytest pytest-asyncio httpx ruff mypy

Process

1. Scaffold — uv init + install dependencies
2. Configure — core/config.py with pydantic-settings, .env, structured logging
3. Define State — TypedDict with Annotated[list, add_messages] for each agent
4. Build Graph — StateGraph with typed nodes, conditional edges, checkpointing
5. Define Tools — @tool with docstrings, Pydantic input schemas, error handling
6. Add Memory — Checkpointing (PostgresSaver), semantic memory (vector store)
7. Add Guardrails — Input validation, prompt injection detection, output validation
8. Expose API — FastAPI routes for invoke/stream with thread_id propagation
9. Write Tests — Basic invoke, tool usage, iteration limit, error recovery, RAG quality
10. Deploy — Docker multi-stage, gunicorn + uvicorn, health checks, Prometheus

Key Patterns

Conventions & Rules

For package layout, LangGraph rules, and FastAPI integration rules, read reference/agentic-conventions.md

Documentation Sources

Before generating code, consult these sources for current syntax and APIs:

Reference Files

Common Commands

uvicorn src.main:app --reload                          # Run dev server (hot reload)
pytest -q                                              # Run tests (quiet output)
pytest -q --cov=src --cov-report=term-missing          # Tests with coverage
ruff check --fix .                                     # Lint and auto-fix
ruff format .                                          # Format code
mypy src/                                              # Type check

Error Handling

For error handling patterns and code examples, read reference/agentic-error-handling.md

LLM provider errors: Use retry with exponential backoff + fallback model chain. Never let provider errors crash the graph.

Tool execution errors: Wrap all @tool functions in try/except. Return structured error messages the LLM can reason about.

Graph infinite loops: Always include iteration_count in state and check it in the routing function.

Post-Code Review

After writing agentic AI code, dispatch these reviewer agents:

• agentic-ai-reviewer — graph correctness, guardrails, iteration limits, cost efficiency
• security-reviewer — tool input validation, prompt injection defense