The Harness Model v2: why we rebuilt it

16 November 2026, 10 min read

Architects copying carved legs

A team's AI maturity is set by what it repeats, not by what its most advanced AI user can do. And it does not average: the weakest aspect erodes the returns of the rest. We rebuilt the Harness Model around that claim: version 2 places a team aspect by aspect, by the practice the whole team repeats.

Imagine a hypothetical team, created for the purposes of this post, that relies on coding agents every day. One engineer supplies the context the agents need, catches the mistakes they keep repeating and decides when their output is ready. The agents are fast, but the team moves at the pace of that engineer's attention, because every correction lands in one head and stays there. The question that matters about this team is not how good its best user is. It is what the team has made repeatable.

The Q2 update that became a rebuild

We published the first version in April as a "Q1 2026 snapshot", said we expected to revise it quarterly, and pointed to a Q2 2026 update. That update did not come. What came instead is a rebuild to fit reality better: not the next edition, and not final either. It grew from six months of writing harnesses ourselves, from the articles we have published since, and from webinars, a hackathon and other events, along with the conversations at them.

The update became a rebuild because our view of what a stage measures sharpened. The first version read a stage from how humans and agents interact, which a team's most advanced user can set; the new version reads it from the way of working the whole team repeats. The model's core premise remains the same, but the stage numbers don't map directly between versions. If you assessed your team using the first version, revisit the page to reevaluate where it stands.

The map in brief

The page states what the model rests on:

Its thesis: once AI takes part in the work, the craft moves to designing the system that controls the AI, the Harness.

The model keeps five stages and names each by where the harness lives:

  • Stage 1 - Ad Hoc AI
  • Stage 2 - Team AI Practice
  • Stage 3 - Reusable AI Workflows
  • Stage 4 - Governed Agentic Delivery
  • Stage 5 - Self-Improving Harness (reserved as a future horizon)

Twelve aspects replace the first version's ten dimensions, in four clusters:

  • People: Role in a Team, Cognitive Load and Time Allocation, Skills
  • Governance: Control, Decision Making, Risk Mitigation
  • SDLC - Software Development Life Cycle: Plan, Build & Deliver, Run & Improve
  • Information Flow: Knowledge, Feedback, Learning

The page also says why placing a team gives a profile:

The result is a profile, not one grade, because a team is rarely at one stage across all aspects. The profile matters because maturity does not average: the weakest aspect erodes the returns of the others over time, like compounding debt rather than a hard ceiling.

The mechanism we argue for is debt, not a ceiling, and the weakest aspect runs it up. Where the team is strong, work keeps shipping at its own speed, while rework, drift and mistakes nobody catches eat into the value of everything shipped. We argued this in public before the rebuild, as "the weakest dimension dominates". Placing your own team is a job for the page.

A stage is where the team's harness lives

In the first version's own words, each stage described "a dominant interaction paradigm between humans and AI agents". The page defines it now:

A stage names where the team's harness lives: what the team has made shared and repeatable around AI. The way of working the team repeats sets it, never one person's most advanced use, and it is not a score to optimise.

By that definition, the opening team's best engineer does not set the team's stage; what the rest of the team repeats does.

The difference starts at the lowest stage: it used to be "No AI Process", while Stage 1 - Ad Hoc AI now means no harness, not no AI.

The strongest objection to reading a stage this way: while a practice is still being invented, the most advanced AI user is the harness in formation, so let them explore. Exploring matters, but a finding becomes maturity only when it lands in the harness the team repeats. Agents free time the team used to spend on work they now do, and that time pays for putting the finding there.

A second objection: there is no reliable criterion for the balance between exploring and the framework a team works within, so any model of what a team repeats will be read as a target. We partly concede it. The balance stays hard to find, and saying a stage is not a score to optimise does not stop a team from chasing one. Only the team's own goals, and how it measures them, can.

Our claim holds for a healthy team of professionals. Where only a fraction of the team cares to advance and the rest keep the status quo and a low profile, it may not hold; we call that a deviation from a healthy team.

If your most experienced AI user were away for a month, which practices would the rest of the team still repeat?

The harness reaches past the code

The thesis covers the whole life of a product, not only the hours spent on code. The model splits that life into Plan, Build & Deliver and Run & Improve, and treats them as a cycle rather than a waterfall: a build can send work back to planning, and what production teaches feeds the next plan. The first version already had rows for planning and for operations, so the rebuild makes the scope explicit rather than discovering it.

When agents make coding cheap, planning takes a larger share of the effort up front, and the discipline moves into requirements, architecture documents and decision records. The page names the principle Discipline Relocates, It Does Not Disappear; the first version opened on the same idea: "The discipline does not disappear - it relocates."

Build & Deliver merges two former dimensions, Workflow & Process and Verification & Quality, because how people and agents share the work of building a change, and how the team judges that it is correct, are two sides of getting one change safely to production.

This is the weakest-aspect argument at the scale of the lifecycle. A harness that stops at the code leaves Plan and Run & Improve behind, and where one of them is the weakest aspect, that is where the coding gains leak away.

If implementation became twice as fast tomorrow, what would still set the pace?

The work that stays with people

Three aspects are new: Cognitive Load and Time Allocation, Skills and Learning. The harness moves load; it does not remove it. Generating a change now costs little, while checking it, fitting it into the system around it, making the decisions and keeping the context current still take people's time and attention, and skills the team has to build and pass on. Our essay on Team Topologies in the AI era argued this in public and called itself "our first input to its next version". Agents free the time producing a change took, and part of it goes into automated checks that report to the agent before anyone reviews the change. The checking that comes back to people is part of the harness, and the harness takes some of it back. Where one of the three new aspects is the team's weakest, it keeps costing the rest, however cheap generation gets.

The same essay asked the next version to "hold the headcount open, assert the load relocation". This version does. Role in a Team describes how the work divides between people and agents, and leaves open whether a team gets smaller or keeps its size and owns more scope.

This is where our answer to the first objection starts to cost something. The time agents free is the first thing leadership wants to bank, as more output or fewer people. Putting it into the harness looks, from outside, like the team got no faster.

What work has AI sent back to people, and have you made time to do it and to teach it?

Delegation runs on the Steering Loop

Control merges two former dimensions, Human-Agent Interaction and Architectural Governance, because both answer where autonomy lies: what an agent may do without a person, where the person engages, and whether boundaries and taste are held by review or by tooling. Feedback separates from Knowledge; the two used to share one dimension. Where the team's knowledge lives is one question; how fast a result comes back and changes the harness is another.

Stage 4 - Governed Agentic Delivery is a change of interface rather than of the AI model. We first argued that in our essay on the agent harness and the work harness, in the first version's stage terms: the workflow is versioned in the repository, and people trigger it instead of prompting it. When a mistake recurs, a person improves the harness rather than correcting the output again. At that stage, changes to the harness are approved by a named team or named people. Autonomy grows only as far as verification earns it, the principle the page calls Maturity Is Trust Earned. So what the team repeats changes through the Steering Loop, not in one person's head.

A third objection: judge the team by what it ships, because lead time and change failure rate matter more than where the harness lives. Outcomes like these are the test of value, and a team that tracks them is measuring. But they show whether delivery moved, not which practice moved it. Choosing which practice to keep or change from those numbers alone falls back on intuition, and intuition is a poor guide to identifying which practices are worthwhile.

Practice-level feedback is what lets a team diagnose which practice to change: what its Sensors (the checks that observe what an agent produced) report, its regression signals, and one agent iteration compared with the last. That feedback helps the diagnosis; it does not prove the cause.

An illustration constructed for this post, with illustrative hours rather than measured ones: a team's build workflow first lets it finish a task in about four hours on average, as its session metadata might show. The workflow grows heavier, with many more tests, and a task now takes over six hours. Lead time alone would at most show that delivery got slower; the workflow's session times show where the extra two hours went. The team judges those hours not worth the results against its own goals, and the practice it changes is the harness itself. That is judgment over a measure; changing a practice by intuition, with no measure at all, is guessing.

Outcomes remain the test of value. Feedback shows what to change; putting the correction into the harness is what lets the rest of the team repeat it. In the opening team, only that engineer can apply the correction when the same mistake returns.

When an agent repeats a mistake, what changes in the harness, and who approves that change?

A map to navigate by

The model is not final. The landscape keeps changing, and we want to keep up with it. What the model gives a team is a map to navigate by and a starting point for its own conversation about where it stands and where to invest next, not a verdict. The page has a section on using it with your team, and one where you can tell us where the model is wrong.

Which aspect would your team place lowest, and what has it been costing the rest?

If this is a live problem for you

Get in touch →