AI agents can now take a specification, work inside an existing codebase and hand back an implementation before an engineer has typed much at all. In an opinion piece for Ctech, Liad Elidan, CEO of Milestone, a startup that measures AI’s impact on engineering work, argues that this moves human effort rather than removing it: “Code written by AI still needs someone to review it, fix it and maintain it.”
The work shifts toward defining intent, supplying context, deciding what to delegate, reviewing what comes back and judging whether it’s ready to ship. Elidan writes that faster generation pushes pressure downstream to review, testing, integration, architecture, security and rework. A pull request may arrive sooner but need more changes before it’s safe to merge, and senior engineers can end up with more validation work.
Measuring outcomes, not activity
“Coding speed is becoming a weak proxy for engineering impact,” Elidan writes. Early enterprise rollouts tracked licenses, active users, prompts and tool usage, which made sense when adoption was the open question. He separates activity (prompts, tokens, generated lines), output (software that survives review and reaches production) and outcomes (faster delivery, better quality and more efficient engineering spend).
His recommended comparison isn’t AI versus no AI. It’s AI-influenced work measured against an organization’s own baselines for delivery, quality, rework and cost, meaning how similar work was delivered before, how long it took and what happened to quality afterward.
Agents that earn autonomy
Multiple autonomous agents make that harder. One agent might generate tests, another refactor code, another investigate incidents and another modify production infrastructure, each with different costs, permissions and acceptable error rates. Elidan argues an agent whose changes consistently survive review and stay stable may deserve more freedom on that kind of work, while others need tighter supervision or a narrower scope.
Costs change too. A seat license is easy to budget, but agents consuming tokens and invoking models at volume introduce a different cost structure, so leaders need to know what the work cost and whether it improved results.
That’s where Elidan sees the engineer’s role changing most. The engineer increasingly defines the boundaries of delegated work, decides what evidence counts as success and determines when an agent has earned more room to act, which requires visibility into the full lifecycle of AI-generated code, not just whether an agent ran.
WebProNews is an iEntry Publication