One agent a week, each co-built with a Champion on your side who owns it. No workshops, no maturity deck, no change programme. Just eight artefacts that exist on Friday and still work when we leave.
The Eight-Week Artefact Sequence is defined as a transformation structured so that every week ends with one working artefact, owned by a named person on the client team rather than by the consultants who built it. Week 1 locks scope. Week 2 builds the Harness. Weeks 3, 4, 5 and 6 each produce an agent: Product, Design, then Coding twice, first on work it can finish in one pass and then on work it has to break down. Week 7 produces a Knowledge Agent and a product-specific Product Brain. Week 8 produces real tests, observability and handover. The sequence matters because each artefact is the precondition for the next one. You cannot trust a Coding Agent before the Harness exists, and you cannot hand a Coding Agent a project-sized ticket before it has earned trust on small ones.
We published very little over the summer. The reason is unglamorous: we were running these transformations rather than writing about them. Everything below comes out of that work, which is also why it is specific about what fails and in which week.
A workshop produces enthusiasm. A week produces an artefact. Only one of those is still there in November.
In 90+ tech due diligences we have read a lot of AI strategies. The failure mode is almost always the same, and it is not technical. A team runs an offsite, picks tools, writes a policy, and six weeks later nothing in the repository has changed. Nobody owned anything.
So we invert it. Each agent is co-built with a Champion on the client side who owns it from the first hour: up to four Champions, one per agent, or fewer people taking more than one. We type alongside them. They keep the thing. This is the only mechanism we have found that survives our departure.
If nobody on your side can name the owner of an agent, you do not have an agent. You have a demo.
The scope is deliberately narrow: the software and the product. Not finance, not HR, not a company-wide transformation. Narrow scope is what makes eight weeks honest. If you want the wider picture of where a team sits before any of this starts, the 4 levels of AI maturity is the framework we score against.
| Week | Artefact | Who owns it on Friday |
|---|---|---|
| 1 | Locked scope: two use cases, access, conventions | Technical lead |
| 2 | The Harness: repo conventions, playbooks, knowledge base | Technical lead |
| 3 | Product Agent: ideas to tickets with acceptance criteria | Champion 1 |
| 4 | Design Agent: clickable prototype from your design system | Champion 2 |
| 5 | Coding Agent on one-pass work: pull requests for review | Champion 3 |
| 6 | Coding Agent on work it must decompose itself | Champion 3 |
| 7 | Knowledge Agent and the Product Brain | Champion 4 |
| 8 | Test coverage, observability, handover | The whole team |
The audit already happened. A one-week AI Factory Assessment read the codebase, scored the team and produced the fixed quote, so week 1 is not a discovery exercise. It converts that into a plan.
Targeted sessions with one or two technical contacts, not the whole team. Two use cases locked in detail. Access granted. Conventions agreed. Champions named out loud, in front of each other, so ownership is a commitment rather than an intention. The artefact is a document nobody can reinterpret in week 5.
The Harness is the part teams skip and then blame the model for. It is the repository conventions, the playbooks and the knowledge base that becomes the tech team's working brain. Without it, every agent run starts from zero context and produces confident nonsense.
Concretely: how a ticket must be written to be actionable, what a pull request has to contain, where the architectural decisions live, which parts of the codebase are off limits. Dull to build. It is the single highest-leverage week of the eight, and it is the one clients are most tempted to compress.
The Product Agent turns ideas into tickets with acceptance criteria and test steps. That is the whole job, and it is more valuable than it sounds, because ticket quality is the upstream constraint on everything a Coding Agent can do later.
Its Champion co-builds it and owns it from the start. By Friday they are the person who tunes it, not us. We consistently see the same second-order effect: once tickets have real acceptance criteria, the humans write better ones too.
The Design Agent takes a defined feature from the Product Agent and builds an interactive prototype from your existing design system. You click the thing before anyone writes production code.
This is where the expensive misunderstandings get caught. A disagreement about behaviour costs minutes in a prototype and weeks after a release. Across 60+ products shipped since 2013, nothing has reduced rework as reliably as making the argument happen earlier.
The Coding Agent picks up tickets already small enough to finish in one go and opens pull requests for review. Humans review and merge. That boundary does not move during the sprint.
We start here on purpose. The work is reliable, the review burden is low, and the team gets to watch agent output land without drama. Trust is built on a run of boring merges, not on a demo. The shape of this is covered in more depth in AI agents are replacing dev teams.
This is the week that matters, and it is the week most rollouts never reach. The ticket is really a small project. The agent has to plan it, split it into sub-tasks, and work through them with review at each step rather than one review at the end.
Everything built in the first five weeks is load-bearing here. The Harness tells the agent what good looks like. The Product Agent gives it criteria precise enough to decompose. Week 5 gave the team the reflexes to review a plan rather than just a diff. Skip any of those and week 6 fails in a way that looks like the model being bad.
Most rollouts stop at the work an agent can finish in one pass. That is the easy half, and it is the half with the least leverage.
The economics are plain. One-pass work makes a developer faster. Decomposable work changes what a team of five can take on at all. The first is a tooling upgrade. The second is why the phrase AI-native means anything.
The Product Brain is a knowledge graph of the product, connected to every agent so they stop guessing at context. Product-specific, not a company-wide brain. Narrow and accurate beats broad and stale.
The Knowledge Agent is the human way in and out. It asks someone to approve what goes in, and it answers questions about what the product already knows. Nothing enters the brain without a person agreeing to it. The maturity ladder behind this layer is set out in the five levels of Company Brain.
Test coverage on what the agents actually shipped, not just the happy path. Then observability: who can see what every agent did and why, trace a bad pull request back to the ticket that caused it, and catch drift when a model changes underneath you.
That last one is not hypothetical. Models get replaced and behaviour moves. A factory without traces cannot tell a bad prompt from a bad week. Our approach to that is written up in how we ensure quality. Then handover: the Champions run the factory while we watch rather than drive.
Agent infrastructure running in your environment, on your own AI subscription, wired into the repo, tracker and chat tools your team already uses. A Product Brain that is yours. Four people who built an agent each and can explain it without us in the room.
We run the same pattern on our own work. Five-plus AI-native products and teams built this way, with an AI Operating System underneath, is where the sequence comes from. We are not describing a method we read about. Twenty-five years in tech, two exits, and this is the first time the constraint has been clarity rather than capacity.
Every sprint is quoted from an Assessment: one week, €4,000 to €6,000, credited against the sprint if you go ahead. That week is where the fixed price comes from, which is why it is the first step rather than an optional extra. The eight-week sprint starts at €26,000. An ongoing subscription with a fractional Head of AI runs €2,000 to €5,000 a month on a six-month minimum, for agent tuning, evaluation of new models and a monthly review.
Below three engineers the economics do not work yet. If that is you, the Product Studio is usually the better door: we build the product with our own factory instead of installing one in a team that is not there yet.
One artefact per week, each owned by someone on your team. Week 1 locks scope. Week 2 builds the Harness: repository conventions, playbooks and the knowledge base the agents read. Weeks 3 and 4 produce a Product Agent and a Design Agent. Weeks 5 and 6 produce a Coding Agent, first on work it can finish in one pass, then on work it has to break down itself. Week 7 produces a Knowledge Agent and a product-specific Product Brain. Week 8 produces real test coverage, observability and handover.
The person on your team who co-builds one agent with us and owns it from day one. Up to four Champions take one agent each, or fewer people take more than one. When the engagement ends, the people running the factory are the people who built it. An agent with no named owner is a demo, not a capability.
Because week 6 is the first week where the agent has to decompose work rather than execute it. Tickets that are really small projects need the agent to plan, split the work into sub-tasks and carry them through with review at each step. That requires a trusted Harness, a usable Product Brain and a team that has already learned to review agent output. Most rollouts stop after the easy pass, which is also the half with the least leverage.
Yes. The infrastructure runs in your environment, on your own AI subscription, wired into the repo, tracker and chat tools you already use. The Product Brain is yours. Each agent has a named Champion who built it. Agents open pull requests and humans review and merge, so engineering control never leaves your team.
Every sprint is quoted from a one-week Assessment, which costs €4,000 to €6,000 and is credited against the sprint. The eight-week sprint starts at €26,000. An optional subscription with a fractional Head of AI runs €2,000 to €5,000 a month on a six-month minimum. Below three engineers the economics do not work yet.
Above The Clouds runs the AI Transformation Sprint and the Product Studio, and assesses AI maturity for investors in due diligence. If your team has agents but no artefacts, the sequence above is the fix. Get in touch to discuss your company or portfolio.