We track where an engineer's hours go When agents write the code, the engineer's job becomes judgment

Review tools haven't caught up with the job

In a software factory, the engineer's job is judgment. Agents do the volume: the scaffolding, the migrations, and the integration checks. The engineer decides whether the plan is right, whether the change is safe, and whether the codebase is getting easier or harder to work in.

We've written about why that judgment can't be automated. This note is about the practical side: the desk where it happens.

The review tools we've used assume the reader is another implementer. They show a change as a raw diff and leave the person judging it to work out the intent from 2,000 lines.

Teams protect focus for the person writing code: headphones, long blocks, no taps on the shoulder. The person judging code rarely gets that. Every decision starts with digging through history, and that's where the focus goes.

We build each artifact for the person deciding

We make the artifact do the work instead of the reader. Every stage of a task produces something a person can take in at once, starting with an implementation plan they review before any code exists.

The change arrives with its evidence and the one question that needs a person's answer. The code is shown visually, so the engineer reads what it means, not every line.

Decisions live in the artifact too. When an engineer redirects a plan or accepts a change, the ruling is logged where it was made, with its reason. Approving takes one click, with no trip to a ticket system.

We measure where the hours go

Both stages of every task, planning and building, run through one platform. So for the first time, we can see where the hours actually go: planning, watching the agents, and validating.

We don't use the split to watch people. We use it to check whether the factory is spending people on judgment or still spending them on volume.

Our working view is that the split shifts over an engagement.

• Building time falls as the agents take over the mechanical share.
• Validating time holds until agents run the end-to-end tests, and then it falls too.
• Planning grows until it's the biggest share of the week.

You see the same log we do

Clients ask what our people are doing if the agents write the code. We show them the log. The planning trail shows every place an engineer redirected an agent before it built anything.

The hours split shows where your engineers' time went, week by week.