AI harnesses
Building the tooling, guardrails and workflows that make AI coding agents productive.
Have Claude record the demo video of your feature
To get a finished local feature signed off: Claude writes a test scenario, plays it in the browser and records it as a captioned video.
My agentic workflow: worktrees, TDD, reviews and a master
5-6 parallel sessions, one worktree per task, an advanced harness, enforced TDD, 8 review agents and a master session: my current agentic workflow.
My AI review workflow: specialised agents in a loop
Specialised review agents, a skill that reruns them in fresh sessions over several rounds, then a final artifact that presents the work that was done.
Review GitHub PRs in a logical reading order, not alphabetical
GitHub sorts PR files alphabetically. A skill that orders the diff into logical blocks and a Chrome extension that reorders the review: 15 minutes to vibe-code.
Todos and human-todos: staying focused while AI codes
Two markdown files at the repo root: AI notes what it spots, I list what only I can do, and background sessions clear the todos while I stay on task.
Strict TDD with AI agents: slower, pricier, and better
Forcing AI agents into strict TDD gives a better design and sturdier tests, at the cost of speed and money. And beyond the prompt, what about hooks?
A good AI harness lets you use much cheaper models
On a codebase with broken import rules, Opus fixes it alone, Haiku falls for it, until an ESLint rule corrects it for 4 to 5 times less money.
ADRs give AI agents the history your code can't tell
README, CLAUDE.md and AGENTS.md only describe the current state of the code. ADRs add the missing history: why that decision, which alternatives were ruled out.
Reaching 10x in the agentic era starts by going 0.3x
To ship 10x faster with AI agents, you first have to slow down: a solid harness, clean parallelisation and mastered CI/CD before you accelerate.
Orca: one glance at every repo, worktree and agent
An ADE that shows every repo and its worktrees, with the status of each agent at a glance: with 3 to 6 parallel sessions, visibility changes everything.
ALL-CAPS Driven Development: shouting at AI does not work
Shouting in capital letters at an AI to enforce an architecture rule does not work: a deterministic tool can check it for you instead.
As many developer setups as there are developers
IDE, ADE, terminal, keyboard or voice, one or ten agents: ways of coding have exploded with AI. My current setup, and a call to compare yours.
Harness Engineer: the developer job of tomorrow?
When you barely write code by hand, the real guardrail for AI is the harness. A portrait of the Harness Engineer, the ultimate senior of the agentic era.
Zero hand-written code: my harness engineering experiment
I set myself the challenge of writing a program without typing a single line myself, pushing the AI harness to its limit, inspired by an OpenAI experiment.
AI harnesses: a fresh, observable local instance
AGENTS.md and skills are not enough: to verify its work at runtime, an agent must run the complete program locally and observe all of it.
AI harnesses: scripts, the underused deterministic tool
A skill tells the AI how to create files; a script does it for the AI, deterministically. Cheaper and more robust: a concrete example with a git-based CMS.
Why I ask for an ADR before every architecture decision
An ADR fits on one page: context, decision, consequences. Michael Nygard's 2011 idea hasn't aged, and it doubles as a free guide for AI reading the repo.
Opinionated frameworks are an AI harness
AdonisJS is famously opinionated. A talk with Jonathan Serra (AI2H) on why that rigidity gives an AI an extra harness, and an open question.
dependency-cruiser: a computational sensor for your AI agents
dependency-cruiser maps imports and fails CI when a boundary is violated. A sensor for AI agents that only works if the architecture is clear enough to describe.
AI harnesses: guides and sensors, the four boxes to fill
A harness is everything in an agent that is not the model. Two axes, four boxes, and a painful conclusion about the harnessability of a codebase.
Your tests are green, but that proves nothing: mutation testing
Coverage measures lines run through, not lines verified. Mutation testing measures whether tests detect bugs, essential once agents write the tests.
How many of your architecture rules break the build?
With AI, a standardised architecture becomes an asset: every rule a machine can verify is a constraint. The others are only intentions.
Claude Code: a TODOS.md file for off-topic details
A warning, an ignored pattern, a bit of debt mid-feature? One CLAUDE.md instruction has Claude Code note them in a TODOS.md file for later cleanup.
The best AI harness is a codebase with a rigorous architecture
DDD, TDD, hexagonal or in-house conventions: it does not matter. A rigorous codebase is the best harness to get an AI to follow your way of working.
The developer isn't dead: Claude deleted its own .claude folder
Mid-feature, Claude deleted the .claude folder: skills and agents gone. Reading the git diff before pushing saved me. A warning for fully automated pipelines.
The dream all-AI-agent dev pipeline… and its weak spot
Agents that do everything, from specs to production, with no human supervision and 100% on Claude. Until the day Anthropic botches an update.
DDD + Clean Architecture: my take after 3 intense months
Three months of DDD and Clean Architecture applied to the letter on a production codebase: the pros, the cons, and why I would struggle to go without.
Claude Code: check that it really read your CLAUDE.md
A Claude Code trick to check it has read both your personal and project CLAUDE.md: have it greet you by your first name at the start of every conversation.