The complete guide · 11 chapters · about 86 minutes
How to build an AI worker with Claude
An AI worker is software that does a real job in your business: it reads, decides, uses your tools and hands over to a person when it should. This guide takes you from the first API call to a worker running in production, the way I build them for clients.
- Developers building their first agent on the Claude API
- Founders and operators who want to know what goes into one before they hire someone
- Teams moving a chatbot prototype to something customers can rely on
A support worker that looks things up in your systems, books, turns risky requests into tickets for staff, hands over to a person when it should, and has a test set that tells you when a change makes it worse.
The chapters
Read them in order the first time; each one builds on the last. Every chapter stands on its own when you come back to look something up.
- 01What an AI worker is, and when not to build oneThe difference between a chatbot, a workflow and a worker, and the four questions that tell you whether you need one.5 min
- 02Choose how to build itA single call, a workflow or an agent; the four ways to build an agent on Claude; and which model to start with.7 min
- 03Set up and make your first callInstall the SDK, sign in without pasting keys into code, and send the first request.5 min
- 04Write its job descriptionThe system prompt as a job description: role, scope, rules it must never break, and when to hand over to a person.7 min
- 05Give it toolsDefine tools as typed Python functions and let the SDK's tool runner do the loop.10 min
- 06Connect your real systems with MCPUse the Model Context Protocol to reach your CRM, database or booking system, read-only first.6 min
- 07Guardrails: approvals, injected text and refusalsWhich actions need a human, why everything the worker reads is data and not orders, and how to handle refusals.8 min
- 08Memory, long tasks and costPrompt caching, effort, compaction, the memory tool and batches: keeping a worker fast and affordable.8 min
- 09Test it before customers see itBuild a small eval set from real conversations and grade it, so every change is measured.9 min
- 10Run it in productionHosting, retries, logging, monitoring and when to let Anthropic host the agent with Managed Agents.9 min
- 11Worked example: a support worker for a repair shopThe whole thing in one file: tools, job description, approvals and a hand-over to a person.12 min
Rather have one built for you?
This guide is how I build AI workers for clients: connected to their real systems, tested on real conversations, with a person always one step away.