Teams that add AI to their review, writing and coding work are finding that more output does not mean more help. The three pieces today count what people actually kept or waited on, instead of how much the AI produced.
Anago (16 Nov · 01), a developer who writes on the Japanese site Zenn, had built three skills for an AI agent that reviews pull requests on a mobile app and its backend. A skill is a set of written instructions that tells the agent how to do a task. An earlier report on those reviews had counted in different ways over different periods, so it could not say which skill worked best. Anago recounted 1,306 comments from six months under one fixed rule: a comment counted as fixed only if the code changed afterwards. The simplest skill had 68.9% of its comments fixed. The most elaborate one ran 11 to 14 review viewpoints in parallel, raised nearly twice as many comments per pull request, and had 59.2% fixed. Anago now runs only the simple skill, because the elaborate version looked thorough while adding comments that people did not act on.
Martin Fowler (16 Nov · 02), chief scientist at the software consultancy Thoughtworks, described a practice that his colleagues Eric Nagler and Unmesh Joshi had used in large organisations. AI models do not know an organisation's private vocabulary, and familiar words used in unfamiliar ways cause the most confusion. Fowler's answer is to have an AI model draft a glossary from internal documents, emails and chat logs, then have the organisation's own experts correct it. The glossary is kept in version control, refreshed on a schedule, and split by part of the business, because the same word often means different things in different departments. The note is short and reports no results, but it gives teams a concrete way to put expert judgement into what an AI agent reads.
Bharat Sharma (16 Nov · 03), a senior engineering leader in platform and data engineering, wrote in LeadDev, a publication for engineering leaders. A fellow leader had asked Sharma how business outcomes had changed after near-complete adoption of AI coding tools. Sharma compared an initiative coded mostly with AI against a similar one coded without it. Coding had taken about 10% of the total lead time, so nearly doubling its speed shortened lead time by about 4%. Sharma calls this one comparison in one organisation. Sharma advises teams to record where work waits, such as code review, decisions and other teams' backlogs, with a reason for every wait.
In each case a person decides what counts: a code change that followed a comment, a definition an expert has approved, or the reason work stopped. Each also keeps a record that can be recounted or compared over time. One inference, which none of the authors states, follows from this. A team should judge AI by what its people kept from the work and by where the work still waits.
All three pieces are thin evidence. One is a developer's count on two repositories, with 36% of comments not traced to any skill. Another is a practice note with no measured outcome, and the third is a single comparison in an unnamed organisation. A larger team with stricter reviewers might find that the elaborate review catches the rare serious bug that the simple one misses.
A developer recounted six months of AI pull-request reviews and found that the most elaborate review skill had the smallest share of comments acted on, so kept only the simplest.
Martin Fowler describes having an AI model draft a glossary of an organisation's private vocabulary, corrected by its experts, so that AI agents use terms the way staff do.
An engineering leader found that AI roughly doubled coding speed on one initiative but cut its lead time by only 4%, because coding was a tenth of the time.


