What an agentic software factory is, and how to govern one

What an agentic factory is not

“Agent” gets used for almost everything. Gartner estimates that of the thousands of vendors calling themselves agentic, only about 130 really are; it calls the rest agent washing: the same chatbots and automation under a new name. It also predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to costs, unclear value or weak risk controls.

So it helps to start with what it is not:

• It is not a coding assistant. An assistant helps one person write faster. The factory runs the whole cycle, from the market signal to the metric in production.
• It is not “let the AI do everything.” Decisions about what to build, how much to spend and when to ship always have a named owner.
• It is not a lab pilot. What leaves the factory reaches production, with a rollback plan and metrics from day one.

Source: Gartner, Jun 25, 2025.

The five parts

An agentic factory is not a tool; it is a set of parts that work together. Remove one and it stops being a factory and goes back to being an experiment.

• 1. Agents with a clear roleEach agent has one responsibility and only the permissions it needs: market, voice of the customer, product, engineering, analytics. Nobody “does everything.”
• 2. The requirementAll work comes in as a requirement: a one-page spec with its user story and success criteria. If it can’t be measured, it doesn’t get built.
• 3. GatesFixed reviews between the idea and production. Agents run some; the ones that commit scope, money or production are decided by a person.
• 4. A guardian agentAn independent agent (Gartner’s term) reviews before every release. If it says STOP, nothing ships.
• 5. Log and metricsEvery merge, release and failure is recorded. Delivery metrics (DORA) are calculated from that log automatically, with no hand-kept spreadsheets.

5 reviews before production

This is the difference between a governed factory and a pile of loose agents. In ours, nothing reaches production without passing these five reviews, and a person makes two of the calls.

1. Idea approved (G0) Man in the loopA person approves the one-page spec: what gets built and how it is measured.
2. Tests first (G1)Tests are written before the code, from the acceptance criteria.
3. Automated review (G2)Quality, architecture, UX, accessibility, privacy and security.
4. Guardian agent reviewAn independent review before release. On STOP, nothing ships.
5. Production approval (G4) Man in the loopA person approves the move to production, with a rollback plan.

Automated gates do the most to lower failure: blocking before merge works better than adding one more human review at the end. The person decides less often, but decides what matters.

What it looks like in numbers: the SERENA case

SERENA is our own corporate wellbeing product, and we built it with the factory. We publish its numbers as they are, including the ones that aren’t good yet.

• 27production releases in 3 days (Nov 6–8, 2026), each with a rollback plan
• 0.45 hmedian time from a change to its merge into main
• 20.6merges to main per day
• 17merges for every human decision
• 25%change failure rate, down from 40%; the goal is 20% or lower
• 39%rework; the goal is 20% or lower

The honest reading: on speed, the factory is already in DORA’s elite group. On stability, not yet: a 25% change failure rate is worse than the high group (20%). It is the same pattern DORA 2025 describes: AI raises the pace and the instability with it. That is why the next investment isn’t more speed, it’s gates that block earlier.

Source: SERENA repository log and metrics; baseline Nov 3–7, 2026 and releases Nov 6–8. Change failure = hotfix merges ÷ production releases. See the full case and its method →

Versus a traditional team and a coding assistant

When it makes sense, and when it doesn’t

It makes sense if…

• Your company already uses AI, but without a system: each team experiments on its own and nobody can measure the return.
• You have processes or digital products that change often and a backlog that never shrinks.
• One person is willing to decide every week what comes next and to check whether it worked.

Not yet if…

• Nobody can own the decisions. Without that person, agents just produce more work with no direction.
• There is no way to measure the outcome. Without a success metric, there is no way to know if a delivery helped.
• Sensitive data has no perimeter: deciding who sees what comes before the agents.

How to start in three steps

1. Self-assessment (3 minutes). Find your starting point across the 7 governance pillars.
2. Agentic diagnostic (2 weeks, fixed price). Agents research, rank and cost your options; you pick one in a one-page Decision Brief. Man in the loop
3. Pilot (4 to 6 weeks). We build the option you chose in production and measure the result against what was expected. Man in the loop

That is why we are called Horme: from the Greek hormē, the drive that, guided by reason, moves toward its purpose. The agents are the drive; a person sets the direction.

The full cycle and the 7 governance pillars, with their real status, are in Agentic governance.