Thinking

Notes from the lab.

Essays, opinions, and technical notes on AI agents, MCP servers, LLM engineering, and shipping software that works.

  1. Guide

    Build your own briefing agent.

    Wire your calendar, inbox, and Slack to the Claude Agent SDK over MCP, keep it read-only, and ship a daily brief you actually trust.

    Read →
  2. Technical

    How we set up and configure Hermes.

    How we set up Hermes — a briefing agent over a founder's calendar, inbox, and signals: the config decisions, MCP tool scopes, evals, and deploy.

    Read →
  3. Technical

    Running MCP servers in production.

    Taking an MCP server from working to production: OAuth scopes, secret isolation, idempotency, versioning, and tracing.

    Read →
  4. Engineering

    Multi-model routing that survives an outage.

    A router that picks the right model per task and reroutes on 429, 529, or refusal mid-request — so one bad provider minute never reaches the user.

    Read →
  5. Engineering

    Evals that catch regressions before users do.

    Build the golden set before the prompt, gate deploys on evals, and feed production traces back so regressions get caught before users do.

    Read →
  6. Engineering

    Observability for agents in production.

    Instrument production agents: loop-shaped traces, per-step and per-tenant cost attribution, alerts that matter, and PII redaction at the boundary.

    Read →
  7. Technical

    Building your first MCP server.

    A practical guide to building a Model Context Protocol server that gives Claude typed, auditable access to your APIs and databases.

    Read →
  8. Guide

    How to choose between Claude Code, Codex, and Cursor in 2026.

    Three different tools, three different theories of what AI-assisted coding should be. Here is how to pick the one that fits your team.

    Read →
  9. Tools

    Claude Code vs. the rest: a working comparison.

    We have used Claude Code, Cursor, GitHub Copilot, and Codex in production. Here is what we actually think.

    Read →
  10. Engineering

    Evals before vibes.

    How do you know your AI feature is working? Most teams go by feeling. The right answer is a small eval set that catches regressions before they ship.

    Read →
  11. Engineering

    Prompts need version control.

    A prompt that worked yesterday can break tomorrow. Without history, you cannot know what changed, what regressed, or how to roll back.

    Read →
  12. Philosophy

    Build the thing you wish existed.

    We run our own products first, then take client work with fewer assumptions and sharper opinions.

    Read →
  13. Process

    Ship the workflow before the pitch.

    The fastest way to lose a technical client is to demo something that doesn't exist. Ship the workflow as a working prototype first.

    Read →

Have something to build?

Start a project → See the work →