Coding with AI Agents

Get reliable, reviewable work out of AI coding assistants and agents, and know where they fail.

For
Developers using AI coding assistants and agents in daily work
Assumes
You write code professionally. No machine learning background needed.
10 steps in 4 stages About 2 h 15 min of reading

Steps open in a new tab, so this page stays where you left it. Steps you've opened turn grey.

Stage 1 · Foundations

What the model is, and what it changes

The mechanics behind everything an assistant does, and the argument for where your value moves.

  1. Guide 24 min
    Core AI Concepts

    Tokens, context windows, and sampling explain most of the "why did it do that" moments in every later step.

  2. Essay 6 min
    AI in Practice: The Skill Inversion

    The argument for the whole path: when code is cheap, judgment is what's left to be good at. The rest of the path is how to exercise it.

Go deeper Core AI Concepts Diagrams

Stage 2 · Basics

Working with an assistant

Giving the model what it needs, and delegating on purpose.

  1. Guide 11 min
    Prompt Engineering

    Most bad agent output traces to context the model didn't have. This is how to give it that context on purpose.

  2. Guide 11 min
    AI-Assisted Development

    Where the speedup really comes from, and why it depends on the specifying and verifying you do around the model.

Go deeper Prompt Engineering Technique Selection Guide

Stage 3 · Intermediate

Real projects, real boundaries

Discipline on a production rebuild, and the rules for what your tools may touch.

  1. Case study 13 min
    When Discipline Makes AI a Force Multiplier

    The basics on a real rebuild with no docs, no tests, and zero tolerance for regressions, including what went wrong before the discipline arrived.

  2. Resource 12 min
    Assumptions-First Task Planning Skill

    The case study's assumptions-first discipline as a skill you can install. Copy it, or use it as the model for your own.

  3. Guide 19 min
    Model Context Protocol (MCP)

    Connecting your assistant to tools and data means trusting a server. This is what that trust covers, and where it ends.

Go deeper AI Security for Organizations Claude Prose Linter

Stage 4 · Advanced

Beyond one session

Keeping long work, and many agents, on track.

  1. Guide 19 min
    Beads: Durable Work Memory for Coding Agents

    Agents forget everything between sessions. Beads keeps the work outside the context window, so long work survives.

  2. Guide 10 min
    Scaling Generative AI Workflows

    One good output and fifty consistent ones are different problems. This is how to split work across isolated agents without cost or quality collapsing.

  3. Resource 6 min
    AI Batch Generation Pipeline Template

    The previous step as files you can copy: a format reference, a plan file, and an orchestrator prompt.

Go deeper Gas City: Orchestrating Fleets of Coding Agents AI Agents Tool Calling