Skip to content

Start here

Everything I know about shipping agents, written down.

Nine guides, 41 minutes end to end. Not a tour of features. The parts that decide whether an agent system survives contact with a real company: how to shape tools a model can drive, what memory looks like six months in, where the human has to stay, and what the security review is going to ask.

These are maintained. When the thinking changes, the guide changes. The essays are the dated version: what broke, when, and what it cost.

01 / Build

Build

The primitives. How an agent reaches a real system, remembers what it did, and gets work routed to the right place.

  1. Designing an MCP server a model can actually drive

    5 min

    Most MCP servers are a REST API with a new hat. The model cannot drive them. This is what changes when you design the tool surface for a reader that has no memory, no screen and one shot at picking the right call.

  2. Agent memory: three tiers, plain markdown, no database

    4 min

    Memory is the part most agent systems get wrong. Not because storing text is hard, but because everything gets written and nothing gets removed. Here is the three-tier split that keeps memory useful past month one.

  3. Orchestrating work with Claude Code: skills, routing and subagents

    4 min

    One agent doing everything does none of it well. This is how work gets routed to the right place, how a repeatable playbook becomes a command, and when a subagent earns its cost.

  4. Claude Code skills and plugins: turning a playbook into a command

    5 min

    A skill is not documentation. It is the judgement you would give a new hire on their first day, written down once so nobody has to give it again. Here is what separates one that gets used from one that gets ignored.

02 / Ship

Ship

What it takes to run inside someone else's company. Identity, the approval gate, and the security review that decides whether any of it goes live.

  1. Auth for agents: the half nobody demos

    4 min

    Every agent demo runs as a god user with an API key in an environment variable. That is fine until it touches someone else's data. This is what has to be true before it can.

  2. Approval gates: designing the place a human has to say yes

    5 min

    Leverage without control is just a faster way to break things. Three categories always wait for a person - and the gate has to be in the system, not the prompt, or it is not a gate.

  3. Enterprise deployment: SSO, SCIM and surviving the security review

    5 min

    The security review is where most AI tooling dies, and it dies over questions that have boring answers. Know them in advance and the whole thing takes a fortnight instead of a quarter.

03 / Operate

Operate

After it is live. Getting a team past the three-enthusiasts stage, and working out why the output was wrong.

  1. Rolling Claude Code out to a team without it stalling at three people

    5 min

    Every rollout gets three enthusiasts in week one and stops. The problem is never the tool. It is that nobody wrote down what good looks like, so everyone else is guessing in public.

  2. When the agent gets it wrong: a debugging method

    4 min

    Almost every bad output traces to one of four causes, and only one of them is the model. Work the list in order and most problems resolve in minutes instead of turning into a rewrite of the prompt.