What is agentic engineering?
4 min read
›Contents
The short answer: agentic engineering is building software with AI coding agents that take on real work, such as writing, testing and reviewing changes, while the team's own tests, gates and review rules decide what merges. The word engineering is the point. The agents do the typing; the team's checks decide what is accepted, and those checks go up before the agents do.
The term is used loosely. This is what it means on this site, and what Kelright's Agentic Guild installs.
How it differs from using an assistant#
An assistant in the editor suggests code as a person types. The person reads each suggestion, and the person is the check.
An agent takes a task, such as an issue or a described change, and works on it in your repository with some independence. It opens a pull request. Nobody watched it write the code, so the check cannot be the person typing. It has to be something that is there every time: the tests, the continuous integration pipeline, the review rules.
That is the shift. With an assistant, quality rides on attention. With agents, it has to ride on the system around them.
What changes in the work#
Four things move from habit to written rule:
- The tests are the specification. An agent can only be held to what is checked. A test suite you can lean on, with a coverage floor enforced in CI, is what lets you accept a change you did not watch being made.
- Conventions are written down for the agent. Agent instruction files carry your conventions in the repository, where the agent reads them as it works, instead of living in people's heads.
- Gates replace vigilance. CI gates hold agent-written changes to your tests before they merge: a coverage floor, a required test on an agent-written change, a limit on how large a change may be.
- Review rules match. Branch protection and review rules say who must look at what, so review effort goes where the risk is.
Together these are guardrails: software that tells engineers and agents what to do at the moment they do it, and that keeps running when the people who installed it leave.
What has to be true first#
You do not start by giving agents more work. You start by making sure the checks exist. The Agentic Readiness Check sorts a team into four phases:
- Fundamentals first. The basics are still going in: tests, a safe way to ship, numbers. The way up is tests you can lean on, and numbers everyone can see.
- A sound house, ready for agents. The foundations are sound and agents are not yet at work on them. The way up is to put agents on real work, behind the checks you already have.
- Agents in use, a couple of things behind. Agents do real work, and one or two checks have not kept up. The way up is to fix those before more of the work moves to agents.
- Already compounding. Agents carry real work, checked before it merges. The job is to keep it that way as you grow, with a record of what you decided.
The failure mode: output outrunning the checks#
There is a fifth place a team can land, and it is not a step on the way up. When agents go up before the checks do, more code is produced than the tests, gates and reviewers can verify, and review is where the work piles up. This is output outrunning the checks. The fix is to tighten the gates, then hand agents more.
It is also why measurement matters. A team in this state can look busy and productive on any activity count. It only shows up when you read delivery before and after, which is what the Benchmark is for.
How a team gets there#
The Agentic Guild installs the practice in thirteen weeks at a fixed scope. It has seven milestones, M0 to M6, and each ends on a check in your own systems, not on a slide: coverage reported in CI, pipeline time, a closed register of tech debt, the share of pull requests agents helped write. It starts with a two-week diagnostic that measures where you are and freezes the baseline, so the change can be measured afterwards. The milestones, one by one.
How to tell where your team stands#
The quickest way is the readiness check: fourteen questions, about six minutes, no account. You get a written read of which phase your team is in and the three places to look first. It is a read, not a score.