explainai

A self-paced course · Coming soon

You've got one agent working. Here's how to run ten.

Most courses stop once one coding agent works. This one starts there: proof instead of claims, work that splits cleanly, review when nobody reads every line, and the numbers that tell you it's working.

See the outline Who it's for

Any agent · Any language · 11 sections · 67 lessons

Ten climbers on ten routes up one wall, watched by one engineer with a tablet

Who it's for

Working engineers who already use one coding agent every day, in any tool and any language.

Before
After
One agent at a time, and you read every diff
Five to ten agents in parallel on one codebase
Done means the agent said so
Every change arrives with proof you can check in a minute
You can't say what the agents cost or produce
You know cost per ticket, yield and where you're the bottleneck

What it isn't: a prompting course, a course about one tool, or a language course.

What you'll learn

  • A single engineer at a desk with a queue of finished work stacking up in front of them

    The bottleneck moves to you

    Once agents are fast, you are the slow part. You'll learn to spend your attention where it changes the outcome.

  • A checklist stamped with evidence: a test result, a screenshot and a number

    Proof, not claims

    Agents say fixed when it isn't. You'll make every change arrive with evidence: a failing test that now passes, a screenshot, a number.

  • A short handoff note passed from one climber to the next along a route

    Context is the ceiling

    An agent is only as good as what it knows. One session per stage, short handoff briefs, and memory you can trust.

  • One wall divided into separate lanes with a climber on each, none crossing

    Work that splits cleanly

    Ten agents on one tangled task make ten copies of the same mess. You'll slice work so agents never collide.

The course project

Every exercise happens in Belay, a booking system for Northwall Climbing, an invented chain with gyms in London and Johannesburg. Members book classes, join waitlists, cancel and get refunds. It's small enough for one agent on day one and big enough for ten by the end.

The spec and the acceptance tests don't care what language you use. Starters ship in TypeScript, .NET and Python.

Bugs it hides

  • Two members book the last spot at once
  • A London class on the weekend the clocks change
  • A payment retry that charges twice
  • A member who can cancel someone else's booking

How it works

  • explainer

    One idea, one picture, no task.

  • exercise

    A real ticket from the backlog, then the solution.

  • walkthrough

    A real session, unscripted, cut down with notes.

  • checkpoint

    Each section ends with a test run and your own numbers.

The outline

11 sections, 67 lessons.

  1. 00 Before We Start 6 lessons Free
  2. 01 Concepts 8 lessons Free
  3. 02 A Codebase Many Agents Can Work In 6 lessons
  4. 03 Proof, Not Claims 8 lessons
  5. 04 Context Is The Ceiling 6 lessons
  6. 05 Slicing Parallel Work 8 lessons
  7. 06 Roles And Model Tiers 6 lessons
  8. 07 Review When Nobody Reads Every Line 6 lessons
  9. 08 Agents On A Schedule 4 lessons
  10. 09 Measure Outcomes, Not Output 5 lessons
  11. 10 Run Ten 4 lessons

See every lesson

Pieter Germishuys

Your instructor

Pieter Germishuys

Pieter builds developer tools and has run coding agents in parallel every day for the past six months: 453 merged pull requests in 176 days, and plenty of failed experiments. This course is what he learned the hard way, including that pull request count measures effort, not progress.

Coming soon

The first two sections will be free. We're building the course now, starting with the lesson Your Agent Says It's Fixed. Prove It.

Read the outline