As AI agents write more code and tests, teams can no longer check everything with the same care. Seven accounts show teams deciding in writing what gets checked closely, by whom or by what, and how they will know whether the work paid off.
Andrew Zigler (14 Nov · 01) writes Dev Interrupted, the publication of LinearB, a company that sells engineering metrics software. He asked engineering leaders where pull request review had stopped being a real inspection. LinearB's data from more than 8.1 million pull requests shows AI-generated pull requests waiting 5.3 times longer for a reviewer. The leaders now sort changes by what they could break, and Postman, an API platform, lets tests and documentation edits merge without a person. Harness, a software delivery company, pauses automatic merging for a class of change after an incident until the missed check is fixed.
Jeff An (14 Nov · 02) is chief technology officer of Momentic, a company that sells AI software testing, and leads eight engineers who use AI agents. His engineers found that agents often declared a feature finished when basic functions failed. A human review of 1,000 unit tests written by AI found only 9.6% useful. Momentic now gives a separate testing agent the original requirements and the running app, with none of the coding agent's history. This is a vendor's account, and part of it describes its own product.
Keisuke (14 Nov · 03), a quality assurance engineer at PIVOT, a Japanese business video company, wrote in Japanese about test cases that AI produced faster than people could run them. PIVOT writes its test cases with AI from product documents, designs and code changes, and one release had 277 cases to run by hand. He now rates each acceptance criterion by how badly a failure would hurt a viewer and whether the release changes it. For one large restructuring release, the cases to run fell from 1,339 to 628, and each skipped case is recorded with its reason.
Yixiong Jiang, Chen Chen and Yuhua Li (14 Nov · 04), of Thoughtworks, a technology consultancy, built SpecBench to test whether extra steps in AI-assisted coding pay off. It runs the same coding tasks through different workflows and scores the results independently. In their sample, adding written specifications, planning and task splitting raised correctness from 92.25% to 95%. The same steps nearly doubled the time taken and roughly quadrupled the tokens used. The authors ask whether the extra correctness is worth the extra effort for the task at hand.
Tony Alicea (14 Nov · 05) writes for Nielsen Norman Group, a user experience research and training firm. He asks how research keeps up when coding agents let teams ship faster than users can be studied. He proposes a short research section in each feature's spec. It states the expected outcome, the signs of success and harm, what to track and when to interview people. A coding agent can then build the tracking with the feature. Alicea warns that this names what to measure but does not prove the feature caused any change.
Postdigitalist (14 Nov · 06), a site that publishes research on AI and design, reviewed the evidence that AI agents save design teams time. The clearest result is a Figma trial with 100 participants, where those using Figma Make, Figma's prototyping tool, took roughly 20% less time on tasks with a reference image. The authors note that no source has yet measured the whole process, including review and fixes. DoorDash, they report, sends a designer's code change through engineering approval and then a design review.
Michael Hill (14 Nov · 07), managing editor of LeadDev, a publication for engineering leaders, reports a survey by Vention, a software development company, of 305 technology leaders. It found that 96% of companies use AI in software development but only 55% have a clear strategy for it. LeadDev's own survey found that only 31% of engineering leaders measure AI's impact. Glyn Roberts of Vention advises recording a baseline of delivery metrics first, and involving engineers in writing the rules that AI output must meet.
Each team stopped treating more checking as automatically better and wrote down where checking pays. The leaders Zigler interviewed sort changes by risk, PIVOT sorts test cases by harm, and Thoughtworks and Postdigitalist count the cost of each added step. Momentic adds a check but keeps it independent of the work it judges, and Alicea and Roberts ask teams to decide what to measure before the work ships.
Sorting by risk depends on predicting where failures will come from, and an AI agent's mistakes may not fall where past human mistakes did. Several accounts come from companies that sell tools or services for the practice they describe. Alicea's piece is a proposal, and Hill's rests on surveys rather than one team's results.
Engineering leaders told Andrew Zigler of Dev Interrupted that AI-written pull requests have outgrown careful review, so they now sort changes by risk before deciding who reads them.
Momentic stopped letting coding AI agents test their own work after finding most of their unit tests useless, and gave the job to a separate testing agent.
A quality assurance engineer at PIVOT halved the test cases run before each release by ranking them by harm, after AI-written cases outgrew manual release testing.
Thoughtworks engineers built SpecBench to measure AI coding workflows and found that adding specifications and planning raised correctness slightly while nearly doubling the time taken.
Tony Alicea of Nielsen Norman Group proposes that researchers add a research section to each feature's spec, so coding agents build the measurement along with the feature.
Postdigitalist reviewed the evidence on AI agents in design and found that measured time savings, about 20% in a Figma trial, come from tasks with a defined target.
LeadDev reports a Vention survey in which 96% of companies use AI in software development but only 55% have a clear strategy, and few measure its effect.






