Week 5 of 10

Agent-Ready Codebases

Core ≈ 3 h 30 min3 videos4 criteria

Focuswhat makes a repository agent-ready: structure, docs, tests, and checks; scoring and auditing readiness; common gaps that block agents in real repositories.

You buildA scored readiness audit and the fixes that make a fresh agent productive

Core material≈ 3 h 30 min

  1. 1
    Engineering case study30 min
    Harness engineering: leveraging Codex in an agent-first world
    OpenAI

    Arguably the single most important reading for this week: repository legibility, structured documentation, machine-enforced invariants, and progressive disclosure.

  2. 2
    Practical guide≈ 60 min
    Repository Harnesses for AI Coding Agents
    Adobe

    A free, systematic guide to making a repository navigable and verifiable by agents.

  3. 3
    Essay20 min
    Harness Engineering
    Martin Fowler

    .

  4. 4
    Videos16 min
    Making Codebases Agent-Ready
    Eno Reyes of Factory

    And Building Reliable Agentic Systems (18 min) — mechanical verification, linters, end-to-end tests, interface documentation, planning, grounding, and human oversight.

  5. 5
    Evals primer20 min
    Your AI Product Needs Evals
    Hamel Husain

    Start the evaluation habit here; Week 10 depends on it.

  6. 6
    Measuring readiness25 min
    The ROI of AI: Why You Need Eval Frameworks
    Beyang Liu of Sourcegraph

    Rigorous evaluations, repository context, engineering KPIs, and avoiding productivity theater.

Deeper material

  • Link
    Unlocking the Codex harness
    OpenAI

    .

  • Video
    Context Engineering
    Revisit Dex Horthy's

    Interview, particularly the harness, loop, and software-factory chapters.

  • Link
    Diátaxis

    A documentation structure (tutorials, how-tos, reference, explanation) that agents navigate well.

  • Link
    Context Rot
    Chroma

    Why a lean, well-indexed repository beats dumping everything into context.

Additional video track

Build

Audit your repository on five dimensions: discoverability, setup, feedback speed, architecture enforcement, and recovery from failure. Score each from 1 to 5 with written evidence. Improve it until a fresh clone can be understood, installed, tested, and changed without tribal knowledge.

At minimum, add a one-command setup, a fast deterministic test path, a short architecture map, local conventions near the code they govern, and a machine-enforced architectural boundary. Give the repository to a fresh agent with one issue and no extra coaching; document every place it gets lost. Re-score afterwards.

Done when

0/4