The blog for engineers building AI agents in production.

Long reads with running code, each written from a lab or a client build. Pick a topic, read the post, then build it live on Thursday.

Save a free seat

Every post links the lab that builds it.

    Every post, by topic

    Coding agents

    1. Search Code with Ripgrep in AgentsGive your coding agent a ripgrep search tool with fixed flags, JSON output, capped results and exit code handling, plus the tool spec the model sees.9 min
    2. Build Safe Bash Tools for AgentsBuild a bash tool for a coding agent that is scoped and time-bounded, with an allowlist, a path guard, an output cap, human approval and a sandbox.10 min
    3. Build a Coding Agent LoopA coding agent is a loop around one model call.10 min

    AI agents

    1. Build LangGraph State MachinesBuild a LangGraph state machine that loops, retries and pauses for a human, with typed state, conditional edges, a checkpointer and a step limit.8 min
    2. Multi-Agent Coordination PatternsThe three ways to coordinate AI agents (chain, router, supervisor), when each fits, the typed handoff between them, and when one agent is better.8 min
    3. Compare AI Agent Frameworks in 2026How to choose an AI agent framework in 2026.10 min
    4. Refine Prompts with Self-CritiqueHow self-refine prompting works, with Python code for a draft, critique and refine loop, when self-critique helps or hurts, and Rephrase and Respond.7 min

    RAG

    1. Detect RAG HallucinationsWhy RAG still hallucinates with a grounded prompt, and how to catch it.8 min
    2. Tune RAG Retriever Top-KWhat top-k means in RAG, why k=5 is a guess, and a 3-step way to tune it with recall@k, a reranker and a holdout set.9 min
    3. Use FAISS for RAG in ProductionIs FAISS a vector database?8 min
    4. Summarize with LangChain Chain TypesLangChain map reduce, stuff and refine summarization compared for 2026.8 min
    5. Build Multi-Hop RAG with LangGraphHow multi-hop RAG answers questions one retrieval cannot.10 min
    6. Choose a Vector DatabaseChroma, Qdrant, Weaviate, pgvector, Pinecone and Vespa compared for RAG in 2026.10 min
    7. Compare Dify, RAGFlow and OnyxDify vs RAGFlow vs Onyx compared for 2026.9 min
    8. Choose a RAG FrameworkLlamaIndex vs LangChain vs Haystack for RAG in 2026.8 min

    MCP and tools

    1. Build LLM Tool Calling with PythonWhat tool calling is, how the request, execute and answer loop works, and Python code for OpenAI and Claude tool calls with Pydantic schemas.9 min

    Production

    1. Retry LLM API Calls with BackoffLLM API retry logic for production agents.9 min
    2. Monitor LLMs with Langfuse and GrafanaPut Langfuse LLM metrics in Grafana with the Metrics API and dashboard JSON, next to Prometheus panels, alerts and trace links.9 min
    3. Build a Chatbot API with LangGraphWire LangGraph into FastAPI with token streaming over SSE, a Postgres checkpointer, per-user thread IDs and human in the loop with interrupt.8 min
    4. Monitor AI Agents with PrometheusAdd Prometheus metrics to a FastAPI agent with prometheus-fastapi-instrumentator, LLM token and latency histograms, multiprocess mode, panels and alerts.9 min
    5. Pool Connections in Python AI ServicesManage database connections in a Python AI agent service: pool sizing, pre-ping, lifespan, PgBouncer, and why no connection is held while the LLM streams.10 min
    6. Persist LangGraph State in PostgresHow LangGraph threads and the Postgres checkpointer keep agent state across restarts and replicas, plus a thread table, per-thread locks and pruning.10 min
    7. Separate Dev and Prod Agent ConfigHow .env, .env.development and .env.production differ, how Pydantic Settings loads them in order, and how to keep agent secrets out of git and laptops.8 min
    8. Run FastAPI Agents with UvicornWhat Uvicorn is, why FastAPI needs it, and the production command, workers, lifespan and SSE streaming setup that keeps AI agents up under real traffic.10 min
    9. Design a Production AI ChatbotAI chatbot system design for production, end to end.14 min
    10. Choose the Right LLM for Each TaskRoute each step of an AI app to the smallest model that passes your evals.8 min
    11. Stream AI Responses with SSEWhen to stream LLM tokens with server-sent events or WebSockets, with FastAPI code for both, plus reconnects, cancellation, backpressure and metrics.8 min

    Read it, then build it live

    Save a free seat