Governed development: how our agents ship work orders through review gates

Agents write code and close work before anyone reads it. Governed development is the discipline that answers for it: scope fixed in a work order before code, separate lanes to build and to check, evidence carried with every hand-off, and a named person who approves what ships.

The short answer

Short answer

Governed development is a way of running AI coding agents so that everything they build can be answered for.

Scope is fixed in a work order before any code, building and checking run as separate lanes, evidence travels with each hand-off, and a named person approves what ships. It is how Mercury's own agents work.

Definition

Governed development

Governed development is the practice of building software with AI agents under a written scope, with separate lanes to build and to check, recorded evidence at every hand-off and a named human approver, so that no agent's own verdict can release its work.

What goes wrong when coding agents ship unreviewed work?

Agents now write code, run their own tests and report the work done, often before anyone has read the change. A green test run or a confident summary gets treated as sign-off. When someone later asks what was checked, by whom and against which scope, the answer is scattered across chat history. And a claim about what the agents achieved travels with nothing underneath it to check.

Security practice already has a name for the underlying risk. The OWASP Top 10 for LLM Applications describes Excessive Agency as the vulnerability that enables damaging actions in response to unexpected, ambiguous or manipulated outputs from an LLM. A coding agent that can merge its own change holds that kind of agency over a codebase, and nobody is placed to stop a bad change before it lands.

Who is governed development for?

It is for technical founders and engineering leads who already let agents write code, and whose agents' output reaches a main branch, a release or a published page. It also answers a buyer's question about any vendor that builds with agents: when the agents built something, who checked it, and who let it through?

How does governed development differ from ungated agent coding?

QuestionUngated agent codingGoverned development
Who sets the scopeThe prompt, as the agent read itA work order whose acceptance criteria are written before any code
Who checks the workOften the agent that wrote it, by running its own testsA separate lane that did not build it
What the reviewer readsThe agent's summary of what it didThe commands each lane ran and what they returned
What a pass meansThe change can mergeOne input to a person's decision
Who lets it shipWhoever presses mergeA named person who read the work, the evidence and the verdicts
What may be claimed afterwardsWhatever the summary saysThe proof step sets the test: a definition, a baseline, a period, a counting method and a source

How it works, step by step

01

Scope is fixed before code

Every piece of work starts as a work order whose acceptance criteria are fixed before an agent writes anything. A build is a plan in the sense the plans and routines page uses. The acceptance criteria are what every later lane checks against.
02

Separate lanes build and check

Gathering sources, writing, building and checking run as separate lanes with separate outputs. The agent that built a piece of work is never the agent that checks it. The separation is between agent instances, each in its own session, so the same model can sit in both lanes. Human teams already have a standard for the check. The NIST Secure Software Development Framework (SSDF) asks teams to perform code review or code analysis, and to record and triage what it finds in the team's workflow or issue tracking system.
03

Evidence travels with the hand-off

Each hand-off carries the commands that were run and what they returned, recorded as produced, so a claim of done arrives with its evidence attached. Software supply-chain security has a word for this kind of record: SLSA defines provenance as an attestation describing how the outputs were produced. Governed development keeps that kind of record between lanes, so a reviewer reads results instead of reassurance.
04

A checking verdict feeds the decision

A checking lane can pass work or block it. Its pass goes to a person as one input, next to the work and its evidence, and carries no authority of its own.
05

The human gate: a named person approves what ships

A named person reads the work, its evidence and the verdicts, then approves the merge or the publication, or sends it back. Nothing ships on an agent's pass. Among OWASP's mitigations for Excessive Agency is to require a human to approve high-impact actions before they are taken. In a repository, GitHub protected branches can require that the most recent push be approved by someone other than the person who pushed it. What a gate is, and where one belongs, is set out on the approval gates page.
06

A proof step for claims

The discipline sets a proof step: a statement of what the practice achieved has to carry a definition, a baseline, a period, a counting method and a source.

Limits

Where does governed development stop?

It does not make code correct

A reviewer can miss a defect. The gate records who approved the change, so a miss traces back to a decision and a name.

It does not hand out autonomy. Whether an agent may ever act without a gate is a separate decision, made on evidence. Shadow mode covers how that is earned.

Some work stays with people

What never goes to an agent is its own question, answered on the when not to automate page.

It carries no result of its own

What the practice achieved is a measurement, and the proof step says what that measurement needs.

Where this fits

Concept

Governed development is a concept page under the AAOS platform, which describes what a work order holds. How It Works covers where the holds sit in an engagement. Two other concept pages explain how work arrives in the first place: the work order you didn't create, and when the system notices your busywork. Every page in the series is listed in the use case library. For business processes rather than code, the readiness checklist tests whether a workflow is ready to hand to an agent at all.

AAOS Platform
How It Works

Questions, answered

Yes. Mercury's agents work under this discipline: written scope, separate lanes, recorded evidence, and a named person's approval before anything ships. What can be shown today is that the practice runs and where its gates sit.
AI Opportunity Audit · Governed development

Find where a person should sign

Bring the place where agent-built work reaches production with nobody's name on it, and find where a person should sign.