Engineering: AI Agents

Six agents sit in the pipeline. Every one of them can be overruled, and none of them can merge, deploy or decide.

The arrangement is not about distrust. It is that generation became cheap while accountability did not, and a pipeline that lets the cheap part authorise itself has removed the only expensive thing in it.

The six agents#

Coding agent#

Writes the change and its tests in the same pass. Tests in the same pass rather than afterwards, because tests written later are written to fit the code that exists rather than the behaviour that was wanted.

Reach: a branch. Never main, never production configuration. Output: a change scoped to one purpose, or a refusal to proceed if the requirement is ambiguous.

Review agent#

Reviews every change, immediately. Its advantage over a human reviewer is not insight, it is that it does not get tired on the fortieth diff of the day, and review quality on the fortieth diff is where defects enter.

It runs before the test suite. A review conducted after a green suite reviews only what the tests failed to notice.

Reach: the diff and the surrounding code, read-only. Output: findings by severity. It cannot approve; it can only fail to object.

Test agent#

Generates cases, fixtures and the edge conditions people skip: empty inputs, boundary values, concurrent access, the second call.

Reach: the test suite and a disposable environment. Output: tests, and an honest statement of what is still untested. The second is the useful half.

Security agent#

Dependency vulnerabilities, committed secrets, and known-bad patterns, on every commit, before CI/CD.

Findings route to Security as well as here, and it does not close its own findings.

Reach: the repository and dependency manifests, read-only. Output: findings with the affected line and a fix, values redacted.

Documentation agent#

Writes from the diff, so documentation is accurate on the day it ships rather than accurate on the day someone last remembered to update it.

Reach: the diff and the docs tree. Output: a documentation change alongside the code change, reviewed with it.

Monitoring agent#

Watches behaviour rather than uptime, and correlates a regression to the change that introduced it. Uptime tells you the process is running. Behaviour tells you whether it is doing the right thing, which is the failure that survives a green status page.

Reach: production telemetry, read-only. Output: a regression signal naming the suspect change.

The five decisions that never move#

1. The architecture decision. Agents generate options and the arguments against them. Choosing is choosing which problems the company lives with for three years.
2. Merging to main. A named person accepts.
3. Any irreversible change. Schema migration, data deletion, public API contract. The test is not confidence, it is whether the mistake can be undone.
4. Declaring an incident and rolling back.
5. Bypassing a gate. Written, attributed, with the risk stated.

Why this holds even as the agents improve#

The list above is not a statement about current capability, so improving capability does not shorten it. Every entry is a place where somebody must be answerable afterwards, and answerability is not a property a model can hold.

That is also the position the first real agentic governance framework arrived at. Singapore's Model AI Governance Framework for Agentic AI, published January 2026, holds that an agent cannot be a principal: the organisation stays responsible, a human stays accountable, and delegation is explicit and bounded. Checked 9 August 2026.

The practical version, for this function: an agent may write anything and decide nothing.

Back to Engineering