Agentic engineering for Python developers

Your agent works in the demo. Then the output drifts, a test passes for the wrong reason, and you can't tell whether the code will hold up. Learn to design the system around the LLM. Start free: a monthly live session, an open Slack and a 12-page guide.

Attend a live event

Free, monthly. Next: 14 Oct.

Join the Slack

Free. Python developers building with AI.

Ship an agent in 6 weeks

Small cohort, 6 seats. Starts October 26, 2026.

150+ developers coached since 2020 · 500+ exercises across Pybites and Rust Platform · 150+ tests per cohort project

Run by Bob Belderbos & Juanjo Expósito

From working code to good design

Real pull request threads from the first cohort: what the tests alone didn't show, and how the code got better, together. Click a screenshot to enlarge it.

Review thread: token prices are per 1M tokens, so the cost estimate is off by a million — A units bug a mocked test won't catch: prices per 1M tokens, used per token

Review thread: the student finds that the provided tests contradict each other, and the coaches update them — The student found that our own tests contradicted each other, so we changed the tests

Review thread: the student corrects the suggested OpenAI type and ty confirms it — It goes both ways: the student corrects our suggested type, and `ty` confirms it

Review thread: a config suggestion the student turns into centralized settings — A week 2 suggestion the student turned into centralized config

Review thread: config moved to typed pydantic-settings with a test for the missing key — Closing the loop: typed `pydantic-settings`, lazy reads, and a test for the missing key

Review thread: a test that passes by accident because Exception stringifies its arguments — A test that passed by accident, and the change that makes it pass for the right reason

Events

Once a month, live. We take one engineering question about AI apps, work through it in code, then answer yours.

See upcoming events →

Community

The Python Agentic AI community on Slack. Ask about the agent you're building, share what you shipped, and get feedback from developers doing the same work.

Join the Slack →

Coaching

Test-Driven Coaching in a 6-week cohort capped at six developers. Pre-written tests, a weekly live call, and a line-by-line review of every PR you push.

See how the cohort works →

Next live event

You know Python and you use AI tools. But how do you build a deterministic system on top of a stochastic LLM?

We build a small agent live and show the engineering around the model:

Tool calling: the agent loop, without a framework
Architecture: a Protocol boundary and a service layer, so the LLM is swappable and mockable in tests
Structured outputs: Pydantic models instead of parsing strings
Human in the loop: the model proposes, your application decides

Then we take your questions. In the last 10 minutes we show how the cohort works, for anyone who wants their own agent code reviewed. Free, and live only.

Go deeper: the cohorts

When you want your own code reviewed. Small groups, a pre-written test suite, and a line-by-line review of every PR you push.

Cohort 1 · starts October 26, 2026 · 6 seats Build a production AI agent in 6 weeks → You build and deploy Expense AI Agent: function calling, structured outputs, a Telegram bot, a web dashboard, 150+ tests, and Docker. Cohort 2 · advanced track Retrieval Engineering → You build askrepo: a coding assistant that indexes a real repository, answers with provenance, and proposes patches that pass the repo's own tests.

From the first cohort

Jeff Haemer, writing software since the early 1980s, on building his agent in the first cohort.

Want to try the teaching style first? Ten browser exercises → · Take the quiz →