Subscribe for updates

Agentic AI architecture needs model routing, telemetry, and policy so production agents can send each workload to the right model instead of one default.

MCP servers should not eagerly load every tool schema into an agent's context window. Lazy-load tools by intent, then govern and audit execution.

MCP packages tools, but the real product is the narrow, typed, auditable interface an agent can actually operate.

Build reliable LangGraph agents with production error handling. Implement retry logic, fallback strategies, and graceful degradation.

Why Focused chose LangGraph as the foundation for enterprise AI agent development. Deterministic workflows, Deep Agents, observability with LangSmith, and eval-driven development in production.

Build evaluation pipelines for LangGraph agents using LangSmith. Learn how to create datasets, run LLM-as-judge evals, and detect regressions before they reach production. Ship agent changes with confidence, not guesswork

Build multi-agent systems in LangGraph using supervisor and swarm patterns. Compare routing accuracy, latency, and real production tradeoffs, with implementation details and failure modes you’ll actually hit.

Build a RAG pipeline that embeds raw PDF bytes with Gemini Embedding 2. No text extraction for embedding, vision-model OCR for the text you need, and better retrieval than extract-then-embed

Our key takeaways from O11yDay NYC on observability, evals, and understanding your systems in production.

Build stateful AI agents with LangGraph memory persistence. Implement checkpointers, long-term memory stores, and cross-session state.

Stream AI agent responses in real-time with LangGraph. Implement low-latency streaming for enterprise applications with five stream modes.

Add human approval gates to LangGraph agents with interrupt() and Command(resume=...), plus the state management bugs that will haunt you at 2 AM.

Cut agent response times in half with parallel sub-agents in LangGraph, plus the failure modes that'll bite you in production.

Improper output handling is one of the most overlooked LLM security risks. This post demonstrates how prompt injection can generate executable payloads and how layered defenses prevent real-world exploits.

You already test. You open the browser, click around, verify the button works. That's a test—it's just trapped in your head. Now watch what happens when you build an AI agent: you tweak a prompt, squint at the output, decide if it's "good enough." That's an eval. The question isn't whether you do it. The question is whether you'll keep doing it by hand.

For the last 10 years, designers have been screaming "it's not about how things look, it's about how things work." The design community got the message. Nobody else did. And with agents, it doesn't matter, because there's no interface to design anymore. The forcing function is gone. What's left is ambiguity, partial information, and systems that need to infer intent from "Book me a flight to Detroit next Tuesday, afternoon, aisle seat." That's still a design problem. We just need a new word for it.