Since December 2025, a growing share of my engineering work has involved supervising coding-agent sessions. The agent can produce a patch quickly. Getting a useful change through a real repository is still mostly a systems problem.

The recurring work is not finding a clever prompt. It is assembling the right context, reproducing the environment, defining what “done” means, checking the result, and deciding what to do when one of those inputs is incomplete.

What fails around the agent

The failures I see most often happen at the boundaries:

The task omits a constraint that exists only in somebody's head.
The repository does not provide a repeatable setup or test path.
A generated change passes a narrow test but breaks a downstream consumer.
One session leaves an artifact that the next session cannot interpret.
A reviewer has to reconstruct why a decision was made.

Better code generation helps with none of those by itself. It can make weak handoffs happen faster.

The pipeline is the useful unit

At Codefac, we are working on agent pipelines that move from a signal or request to an implemented, checked result. A simplified version has four stages:

turn the request into a scoped task with the relevant context;
produce the change in a controlled runtime;
test and review the change against explicit acceptance criteria;
preserve enough evidence for the next person or agent to continue.

Humans still own the specification, permissions, risk decisions, and exceptions. The goal is to stop spending human time reconstructing routine state between stages.

What I would measure

Raw lines of code and the number of concurrent sessions are not useful measures of this system. The measures I care about are:

completed changes after review and rework;
the share of artifacts that pass to the next stage without being rebuilt;
defects that escape the checks;
retries caused by missing context or environment failures;
human time spent on exceptions rather than routine coordination.

These are operating measures for the pipeline we are building, not universal KPIs for every software team. The claim is narrower than “10x engineer” or “100x pipeline”: as code generation gets cheaper, the quality of the handoffs and checks determines how much of that output is actually usable.