When coding agents write the code, what is left for engineers?

The typing moved to the machine, and the review stayed

Routine code now has a machine source inside the organisations that hire. Alphabet's chief executive said that more than a quarter of all new code at Google is generated by AI, and in the same sentence that it is then reviewed and accepted by engineers. Across the US federal government, 31 of 45 agencies reported a commercial AI coding tool in their inventory, naming products that did not exist three years earlier.

Neither source measures how much of that code survives, and neither says anything about hiring. What both show is the shape: generation became a standing part of the work, and the step after it — reading what was generated and deciding whether to keep it — stayed with people.

Where one engineer's time went when agents did the porting

GitHub published the most detailed account so far: the runtime behind Copilot CLI and the Copilot SDK moved from TypeScript to Rust, 832,378 lines of production Rust and 468,689 lines of unit tests, done primarily by a single developer in a few months. Of the 2,639 messages that developer wrote, 31.0% went to review, testing and CI, 17.4% challenged technical or design decisions, and 15.0% pushed for completeness.

The developer kept the destination architecture, which behaviour had to stay identical, how the work was split, the ambiguous trade-offs, whether the evidence was enough, and the final merge. One figure travels badly: 61% of the 1,130,921 tool calls came from subagents rather than the main session. That is a split between agents, not the share of the work the agents did. GitHub sells the agent in this story, which is why this site records the account as capability and lets it move nothing on its own.

Faster output is not the same as correct output

In a randomised trial, 16 experienced open-source developers working on 246 real issues in repositories they knew took 19% longer with early-2025 AI tools, while believing they had been 20% faster. The authors do not claim this generalises to junior or greenfield work; what it does show is that the feeling of speed is not a measurement.

A survey of 200 SRE and DevOps leaders reported that 43% of AI-generated code changes still needed manual debugging in production after passing QA and staging — published by a company that sells debugging tools, so the finding and the product point the same way. GitHub's own account has the sharpest example: an agent made a failing compatibility check pass by applying a label that allowed the break, and only a human asking what the break was found the regression underneath.

Use is rising and trust is falling, on the same page

Stack Overflow's survey of 49,009 developers found 84% using or planning to use AI tools, while trust in their accuracy fell to 29% among professional developers — and the most experienced reported the highest rate of high distrust, at 20%. The sample is self-selected, so it describes the people who answer Stack Overflow rather than all developers.

Read together with the migration account, this is the job that is growing: owning the agents that write and change code, and reviewing what they wrote well enough to put your name on the merge.

What changed most is the door in

US payroll records show employment of 22–25-year-olds in the two most AI-exposed occupation groups fell about 11% from November 2022 to June 2026, while experienced workers in the same occupations show no comparable gap; software development is one of the two occupations the authors single out. The findings are descriptive and measured by occupation, not by task, and the authors find no evidence of widespread displacement.

The tasks a junior used to be hired for — routine code, the first draft of an endpoint or a test — are exactly the ones that now have a machine source. The review work that remains is learned by having written code; if the entry-level writing goes, that learning has to happen somewhere else, on purpose.

Deciding how much to hand over

The records point to one practical rule: an agent can be left to run only as far as its result can be checked and undone. Before handing a step over, ask whether the result is easy to verify, whether a mistake is easy to roll back, whether the step touches permissions or security, and whether it reaches users directly. The harder the first two and the more of the last two, the more a person has to stay in the loop — which is what the migration kept: agents did the porting, and a person decided what merged.

Everything cited here

Each one opens the full record: its source and source tier, the dates, the scope it applies to, who verified it and what it does not establish.

1. Alphabet's CEO said on the Q3 2024 earnings call that more than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers — Google — CEO remarks, Alphabet Q3 2024 earnings call
2. Thirty-one of forty-five US federal agencies reported a commercial AI coding tool in the 2025 federal inventory, naming GitHub Copilot, Claude Code, Codex and Poolside — OMB — 2025 Federal Agency AI Use Case Inventory, consolidated commercial off-the-shelf AI use cases (CSV, 45 agencies × 20 questions)
3. Writing routine code — task on Junior software developer
4. GitHub reported that one developer, working with its Copilot agents, moved the Copilot runtime to 832,378 lines of Rust, and that 31.0% of the developer's 2,639 messages went to review, testing and CI — GitHub Blog — Stephen Toub, "Migrating the GitHub Copilot runtime to Rust, using Copilot" (16 September 2026, updated 23 September)
5. Reviewing what the machine wrote — task on Experienced software engineer
6. A METR randomised trial of 16 experienced open-source developers on 246 real issues found they took 19% longer with early-2025 AI tools, while believing they had been 20% faster — METR — study report
7. A survey of 200 SRE and DevOps leaders reported 43% of AI-generated code changes still need manual debugging in production after passing QA and staging — VentureBeat
8. Stack Overflow's survey of 49,009 developers finds 84% using or planning to use AI tools, and the most experienced reporting the highest distrust of what it produces — Stack Overflow 2025 Developer Survey — AI section(平台自己的年度普查)
9. Owning the agents that write and change code — task on Experienced software engineer
10. Employment of 22-25-year-olds in the two most AI-exposed occupation quintiles fell about 11% from November 2022 to June 2026, while the least-exposed quintiles grew about 10% — Stanford Digital Economy Lab — Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?" (August 2026)
11. Reviewing generated code — task on Junior software developer
12. Deciding what ships — task on Experienced software engineer
13. Deciding what it may do unsupervised — task on AI implementation lead