The Harness Is the Moat

The model is no longer the hard part. The harness is: the vision, rules, gates and records that tell an agent what done means and who decides. I gave a talk on this at the Claude Operators Summit in London on 19 November 2026. This is the written version.
I am not a developer by trade. I started in sysops, moved through cloud engineering, and now work as a Technical Director at Nearform. That matters, because it means I built my harness to cover my own weaknesses. Yours will look different, and it should.
The question has moved
In 2025 we argued about copilots. Which assistant, which vendor, how much it autocompletes. That conversation is over. The models now plan, break down and pursue a goal without hand-holding, and they do it well enough that the vendor stops being the interesting question.
The interesting question moved to the operator: what does done mean, who makes the calls, and how do people and agents work together? A copilot makes you faster at a job you already do. I got tired of the word. If an agent can do the job, why am I still doing it?
One operator, one year of evidence
My main test bed is C4X, a C4-model extension for VS Code. It draws architecture diagrams next to the Markdown that explains them. It needs an instant preview and useful exports, with no server and no Java.
My definition of done for a side project is simple: someone else uses it. It lives in a public extension store, and strangers have installed it. Not "works on my machine". When I gave the talk, C4X had passed 10,000 installs and downloads; the apps page carries the current figure.
The git log tells the story of how my thinking changed. The first day, Sunday 19 October 2025, brought 33 commits. Version 1.0.0 followed 45 days later. In my private dev repo there are now 1,362 commits: 39 in May, 178 in July, 203 in September and 582 so far in November. I am one developer, so that is correlation, not proof.
Early on I chased models. When a new one came out, I waited for it. Then the order flipped. Instructions, hooks and a definition of done were retrofitted between May and July 2026, and the harness became my first priority, ahead of the model.
Build on the shoulders of giants
I did not invent any of this from nothing. BMAD gave me AI personas that run phased workflows, though the ceremony grows with the size of the project. Spec Kit gave me spec, then plan, then tasks, though the spec documents pile up into noise. Superpowers gave me test-driven development, debugging and plan execution, and it is now an optional integration in my charter.
Each gave me a workflow. None told me who decides, or what done means. That gap became DevEx Charter, my open-source plugin for Claude Code, listed on the plugin hub. It sets out who is accountable, what is allowed, and what "the checks pass" means. It is opinionated, and not for everyone. You probably have your own version taking shape on a laptop somewhere.
The golden thread
Every person and every agent session in my projects starts from the same four files:
- ROADMAP.md says what is next, and in what order, with the exit evidence that closes each milestone.
- STATUS.md says where we are now: the current release, the work in flight, the known limits.
- CHANGELOG.md says what shipped, and with what proof.
- OPERATOR.md is the one that is mine. It holds every decision only I can make, each with a recommendation.
A 23-line CLAUDE.md points at them first, and a planning check fails the build when the records drift. I also keep a short vision file, so that when the agent loads up and I ask it to explain the project, we agree.
On top sits a delivery orchestrator, the skill that picks skills. I talk to it every day. It decomposes the work and routes it to experts for specs, architecture and test strategy, and it keeps the records straight. I keep the vision, the funding, the one-way doors and the release call.
From asking for tests to asking for proof
For a long time I ended every prompt with the same line: make sure you have tests, and get them passing. The model said yes, and I took its word for it. I now ask for proof.
Skills suggest. Hooks, tests and gates prove. A QA validator is required on code changes: tests exist, coverage maps to requirements, regression runs. The test report is parsed from an artefact, never typed by the model. And the definition-of-done hook changed its job. From May it only printed advice. Since 7 August it blocks when Claude says it is finished.
The harness lied to me
It lied in three ways, and I have the receipts.
- Green, but never ran. A tag was filtered out of CI. A check passed whatever happened.
- Flagged, but not acted on. Non-deterministic checks found real issues. The model carried on, and the rest stayed green.
- Failing for the wrong reason. Pre-commit rejected every commit over a setup gap, not over the code.
My rule is to make every gate fail once before I trust it to pass. Deterministic checks decide. Non-deterministic ones only advise, unless something is made to act on them. A model that wants to please you will tell you what you want to hear.
Green also missed the user. On 14 October I installed the 1.8.0 candidate and clicked through it. Two faults turned up within minutes: the preview's Edit link did nothing, and a right-click in the second diagram opened the first. The test suite was green throughout. I fixed both by 16 October, re-drove the path and shipped. Install your own build and follow the user's path.
The author never marks its own homework
Before I commit, push or merge large work, a different model tries to break it. Deterministic gates run first. Then a brief asks, "how would I break this?", and the reviewer's job is to refute, not to approve. I use other subscriptions for this, run headless on my machine. Over the year that meant Codex, Kimi and the Antigravity CLI; today Codex does most of it. If none is available, I fall back to the same model in a fresh context window. That is independent, but not diverse.
I call it my AI council. It caught things where one model had been leading me in one direction for too long.
Same harness, next project
What transferred between projects was the definition of done, not the code. ExamCertPrep, a site for preparing for Anthropic and Google certification exams, went from nothing to live between Friday 24 and Sunday 26 July 2026, in 186 commits.
Then I had to rebuild it. When I tried to scale it, the architecture did not hold: the spec was wrong, not the model. Fast and wrong is still wrong. Intent comes first, every time, and the architecture decisions now live in written ADRs. A second weekend got it back on track, and it now has over 100 users. There has been no marketing.
Which manager are you this week?
Once the agents work, the problem becomes managing them. I see four types:
- The Know-It-All delegates, then redoes the work.
- The Micromanager feeds small tasks and reviews every step.
- The Goal-Setter defines outcomes and checks the evidence.
- The Fire-and-Forget delegates and never checks.
The Goal-Setter and the Fire-and-Forget both delegate. Only one checks. I move between all four, depending on the day, but I am trying to live as the Goal-Setter. I wrote the longer argument in Your AI agents have a management problem.
My own use has matured through three layers. First the copilot: one human, one AI, every diff read. Then the team lead: one human, several agents, with outcomes checked rather than every line. Then AI pods, where orchestrators run their own agents and I act as the CTO, setting intent, grading outcomes and making the release call. Each step up needs more from the harness, not more agents.
Agents also do not shorten the week. There is more output, so there is more to decide. I am out of context by Wednesday.
Whose harness is it?
A personal harness captures how I work. A team harness has to agree how we work together. Mine suits me, and my team uses it, though not everyone likes it. At team scale the failure is stale context, not missing context. Someone has to own the authoritative source, the ADR log and the status file. When nobody does, decisions go stale or get made twice.
Monday morning
Five things to try this week:
- Write down what done means. One page, before the first prompt.
- Make one gate fail on purpose, then trust it to pass.
- Install your own build and use it, the way a user would.
- Get a second model to review mid-flight. A different model, not the author.
- Ask which manager you are this week, and whether your harness fits that layer.
Scope, risk and the release call stay with you.
Where to go next
Everything above is a start, not an answer. The projects are open for you to copy: C4X, DevEx Charter and ExamCertPrep. The talk itself is on my speaking page.
For the bigger picture, the question of what production looks like in your organisation, and how to get from zero to something real instead of another proof of concept, read my book, Building the Agentic Enterprise on Google Cloud. It is a field guide to why agentic systems fail after the demo, and how to run them in production. Get it on jpantsjoha.com/book.
Jaroslav Pantsjoha
Technical Director · Agentic AI & Cloud Platforms · Google Developer Expert
Full bio →