AI agents now produce more code and more conversations than people can read one by one. Three pieces published this week show teams deciding what a person still looks at, and what they count as progress.
Gergely Orosz (16 Oct · 01) writes The Pragmatic Engineer, a newsletter for software engineers, and he interviewed Uday Kiran Medisetty, a Distinguished Engineer at Uber, the ride-hailing and delivery company. Uber's automated build and test jobs, which run on every code change, more than doubled in twelve months, and the company declared an internal "code red" for that system. Simple checks such as linting moved to the engineer's own machine, where AI agents run them before opening a pull request. Engineers still review every pull request, but an AI review, screenshots or videos and a risk score arrive first. Uber also stopped relying on how many pull requests reach production and now tags each one as feature, bugfix, maintenance, testing or docs.
Sarah Wells (16 Oct · 02) is an engineering leader and consultant who has just become Editor at Large at LeadDev, a publication for engineering managers. In her first column she names five changes she plans to follow, and two concern how teams work with AI agents. She writes that teams are getting smaller and that engineers increasingly work with their agents instead of with each other. In her view, working together was how context spread, how junior engineers learned and where people challenged assumptions, and she is not sure anything has replaced it. She also cites a Harvard study of 718 firms in which AI agents raised coding activity but not shipped work, partly because code review became the bottleneck.
Teresa Torres (16 Oct · 03) writes Product Talk, a site about product discovery, and she interviewed the team behind Your360 AI. The company offers interview-based 360-degree feedback, in which a voice AI coach called Tam interviews colleagues and talks the recipient through the results. The team includes Jared Goralnick, the co-founder and chief executive, Eda Goksel Jackson, the head of design, and Alex Krasun, an AI engineer. A single prompt could not hold a structured conversation, so they replaced it with a state machine, a set of conversation stages with rules for moving between them. The team also stopped reading routine call transcripts, and automated flags on sentiment, customer value and product quality now decide when a person steps in.
In all three pieces a machine takes over a routine check, and a person keeps a narrower decision. Uber's reviewers see an AI review and a risk score first, and the Your360 AI team reads only the calls that automated flags select. Uber and Wells also question measures that teams used before AI agents, such as pull request counts and coding activity. The link between the three pieces is an inference, and none of the authors draws it.
Uber's account comes from its own engineer on a newsletter podcast, and it gives only the direction of change, not figures, for its new measures. Wells herself says the research on productivity is messy, so the Harvard result may not hold for every team.
Uber's engineers still review every pull request, but an AI review and a risk score come first, and Uber now tags pull requests by type instead of counting them.
Sarah Wells, LeadDev's new Editor at Large, warns that engineers who work with AI agents instead of each other may lose how context spread and how juniors learned.
Your360 AI replaced a single prompt with a state machine, so its voice AI coach keeps a structured feedback conversation on track while still going deep.