The method
The whole view decides the design.
Stand where the whole system is visible before deciding anything. The test of a decision made from there is whether the team still owns and runs what was built when you come back to it, and whether do not build this was ever an available answer. Crinaro is a ridge line, from the Italian crinale: the crest path where you can see down both sides.
Where you are
Agents went in,
your costs went up. Why?
It is not just you. McKinsey asked 1,719 people across 97 nations for its August 2026 report.
say AI is scaling across the enterprise. The only line above that moved.
McKinsey, 1,719 respondents, p. 5 (opens in a new tab)say it reached EBIT. A survey answer, not an audited line.
Of respondents, not of companies, p. 12 (opens in a new tab)attribute 5% or more of EBIT to it. That share has not moved either.
And report significant value. McKinsey's AI high performers, p. 17 (opens in a new tab)of that 6% report redesigning the work itself. One in four of the rest do.
Nearly three-quarters, and n = 92 against 1,429 others, p. 18 (opens in a new tab)Those figures are leaders grading their own organizations, and a second firm reports a similar distance between adoption and return.1 So the question is not whether you are adopting AI. It is whether the way you adopt it can move EBIT.
The AI often changes nothing, McKinsey's advice says (opens in a new tab), when the work on either side of it is left as it was. A model can predict a machine failure days ahead, and maintenance still runs on the calendar. The prediction changed. The way the decision gets made did not.
The way you adopt it is where the problem starts. Every team of people you run already answers five questions: Who owns this? Where do they get their facts? How does a decision get made? Who does the work? Who builds the team?
Nobody adds fifteen people without answering them first. Agent teams are getting stood up without the same controls. Each question has an answer further down, and none of the answers is a tool.
Teams answer “Who owns this?” by adopting AI the way they already work, so it amplifies the organization it lands in and costs continue to rise. DORA found the amplification in AI generally (opens in a new tab), across nearly five thousand practitioners, most of whom were not yet using agents.
The other four stay nobody's job, so the system keeps the shape of the teams that built it: why do we have five user APIs.
The technology is new. The problem it lands in is not: it is flow and ownership, which you already manage everywhere else.
Projects and acquisitions leave the same shape behind: three teams with three answers to the same fact, and agents that read all three and answer with confidence.
Who owns this?
A name against each capability
Not whoever picked it up. One team answerable for each part of the system after the current deadline, and an operating process that keeps it that way.
Where do they get their facts?
One place the facts live
How your system works, and the procedures people once learned by training, written where an agent can read them. Agents decide from what they are given. Without it, they guess. Who owns the answer
How does a decision get made?
The why, captured when it is decided
Code says what it does. It never says why it was built that way, or which duplicate is meant to go. Written down by whoever made the call, at the time, not reconstructed later by somebody who was not there.
Who does the work?
A route from a question to its owner
Not whoever is nearest. The question reaches the team that owns it, and when something of theirs is broken they fix it, rather than everyone routing around it.
Who builds the team?
A team that builds the teams
Agents do not arrive trained: somebody designs them, checks their work and keeps improving them. In engineering that may be the team that runs them; elsewhere it may be a separate group. Either way it is a role the organization stands up and staffs, not work added to the side of someone's desk.
- KPMG, Global AI Pulse, September 2026 (opens in a new tab), p. 7: a second firm, across two waves of its own study. 13% of 2,131 leaders were still experimenting, against 22% in April, while the share reporting an established return changed less, from 8% to 10%. KPMG reads movement between quarters as direction rather than step change.
Where to start
Own what is nobody's job.
If you can maintain the platform you have and keep building on it while you hit your financial goals, keep going. If you cannot, this is where to start.
First, decide who owns what. Only about a third of the 2,145 leaders KPMG asked (opens in a new tab) for its June 2026 report said the roles for running AI day to day were very clear and well managed.2 An executive who owns AI and its controls is the start. Somebody also has to own the design: each process, each system and the data behind it, and how work moves between them.
Then fund an answer to each of the five questions above, and the platform they run on. None of it is a tool purchase. You can buy a product that maps what you already have; you cannot buy somebody answerable for keeping the map true. What the gaps in your machinery cost
Budget for what ownership costs. Your delivery dates will depend on other teams' queues. Teams need time before they take on the change. Removing what is already duplicated is work that has to be funded, and so are the people who build the agent teams. Treating any of it as free is the most expensive mistake here.
The agents are already going in,3 and 74% of the 3,235 leaders Deloitte asked (opens in a new tab), as it reported in April 2026, expected at least moderate use by 2027. Designed first, the system around them can take cost out. Left as it is, they amplify the organization you run today, costs included, and that is a hard result to take to a board.
Worth bookmarking
What you can use, layer by layer. The options at each layer of an agent setup: a list, not a review, and every row carries the date it was read, so each visit shows you how current it is. See the options
- EY, AI Risk and Governance Survey, September 2026 (opens in a new tab): a different firm, on controls. Of about 180 senior AI executives at US companies using agentic AI, 49% said their governance framework had not been updated for it, and 26% said they could not detect unauthorized agents operating internally.
- KPMG's September report again, note 1, not the June one above, p. 3: of 2,131 leaders, 38% said their organization was developing or implementing multi-agent systems, and significant employee adoption of AI agents had risen to 34% from 25% in April.
What success looks like
Fewer places to change,
year over year.
If your agent teams are working, the number of systems you have to change to deliver one thing the business asked for goes down. That is the measure.
The 6% who attribute 5% or more of EBIT to AI changed the work, not just the tools, three times as often as everyone else. Prompt-style use makes a person faster without making the work repeatable. You do not get a different number out of the same system. McKinsey's September 2026 technology outlook (opens in a new tab) puts it the same way: agentic AI is changing operating models, not just tools.
It may simply be early: economists call the lag the productivity J-curve (opens in a new tab), where the gain arrives only after the work around the technology is rebuilt. Either way, somebody has to own the rebuilding.
Success is not more agents in more places. It is a count that goes down year over year, because each thing the business asks for touches fewer systems that somebody has to keep in agreement by hand. Count the repetitive changes, not the repositories
Where this comes from
Patterns learned in hard places.
None of this is industry-specific. The experience behind it is: regulated, legacy-heavy, and full of decisions that are judgment calls rather than lookups. What holds up there travels.
How the work gets done
The model is under test, by being run.
The model this page argues for is not described: it is run, and still changing. Three agent teams keep it under test. One curates AI-SDLC, the reference on delivering with agents that this argument comes from. One maintains this brand and the page you are reading. One runs a public marketplace you can open and install: github.com/crinaro/marketplace (opens in a new tab). Three different problems on purpose, so the approach gets challenged rather than confirmed. The three teams, and what each keeps alive
If any of this is useful
Say where it breaks.
These are patterns, not prescriptions, and the interesting mail is the mail that says a piece of it does not hold: in your architecture, at your size, with the constraints you actually have. That is a conversation worth having whether or not anything follows it.