AI coding agents and what they do to a software team

Two years ago AI in the editor meant completing the line you were typing. Now a developer can hand an agent a ticket and get back a branch with a working change, tests and a description. On well-structured codebases with good tests, this works often enough to change how a day feels.

The effect on a team is not simply faster. It moves the bottleneck.

The generation is cheap, the review is not

When producing a change costs almost nothing, more changes get produced, and every one still needs a human to understand it, judge it and take responsibility for it. Review becomes the constraint. Teams that win here get ruthless about small pull requests, invest in fast and trustworthy CI, and treat a reviewable diff as the unit of work rather than a feature.

The last 30% is still the hard part

Agents are strong on the well-trodden 70%: CRUD, wiring, refactors, tests for existing behaviour, migrations. They struggle with the parts that need context that is not in the repo: why this constraint exists, what the product actually needs, the non-obvious performance cliff, the security implication. That 30% is where senior engineers now spend their time, and it is where the value is.

What changes for how you staff and plan

• Fewer people can maintain more surface area, so team size stops tracking scope so tightly
• Junior work that was learning-by-doing is now partly automated, so mentoring has to be deliberate
• Estimates compress on routine work and barely move on genuinely novel work
• Test suites and CI go from good practice to load-bearing, because they are what makes agent output safe to merge

Our take

Used well, coding agents let a small senior team cover the ground of a larger one, which is a model we already believe in. They do not remove the need for engineers who understand the system. If anything they raise the premium on that understanding, because someone has to be accountable for everything the agent writes.