What an agent actually is (and what it isn't) Free
Tool calling: schemas, validation, side-effect safety
Build: a one-shot agent that reads your inbox
A four-week, hands-on workshop for engineers and operators. Wire LLMs to real tools, design loops that recover from failure, and put agents into production behind your own MCP server.
This isn't a survey of frameworks. By the last session you'll have a running agent in your stack and the judgment to keep building more.
Wired to your tools, with retry, budgets, and a recoverable loop. Deployed to staging by week 3.
Hosted, scoped, and auth'd — so any AI client can invoke your domain logic on a user's behalf.
When to plan, when to act, when to stop. Where loops break in production and how to keep them honest.
Sessions are 90 minutes — half lecture, half pairing. Recordings up the same evening. Each module ends with a build to merge before the next one.
You're comfortable in TypeScript or Python and you've called an LLM API before. You don't need to have built an agent — that's what the four weeks are for.
Move past notebook demos. Learn the loop, retry, eval, and deploy patterns that hold up when real users push real input through your agent.
Build enough to know what an agent can and can't do this year. Leave with a prototype you can hand to your team and a vocabulary that matches reality.
Mira's spent the last three years putting LLM-driven systems behind real products — first on internal tools at Linear, then on agent infrastructure at Anthropic. She writes about the parts that aren't on Twitter: the failure modes, the eval pipelines, the boring code that makes loops survive Monday morning.
Plan on 3–4 hours of building between sessions. Every module ships with a starter repo, a target output, and a reference implementation you can read after you've tried it yourself.
We default to Anthropic's Claude and OpenAI's GPT-4 family because the tool-calling story is mature, but the patterns generalize. You'll get a $40 credit on each provider for the duration of the cohort.
No. Week 03 walks through the protocol from first principles — what it solves, why scoped OAuth matters, and how to host your own server. If you already know it, you'll skip ahead and help review pull requests.
Yes — full refund within 14 days, no questions. After that we'll prorate against sessions attended.
Recordings post the same evening with chapter markers. The community channel stays open between cohorts, so you can ask questions long after demo day.