GM's autonomous vehicle engineers spend 15% of their time writing code. Rashed Haq, GM's VP of autonomous vehicles, disclosed that figure at VB Transform 2026 and said AI agents now handle large portions of the remaining 85%: analyzing vehicle telemetry, triaging problems, running machine-learning experiments, and testing fixes. The outcome is three times as many merged pull requests, faster feature releases, and fewer defects escaping into later development stages.
The method matters more than the result. GM did not deploy coding assistants. It mapped each engineering loop, simulation, public road testing, and post-deployment monitoring, identified the longest bottleneck in each, and automated it. Agents connect to internal tools and petabytes of vehicle data through customized Model Context Protocol servers. Permissions mirror the individual engineer's access rights, and engineers remain accountable for agent output. GM also version-controlled instruction documents it calls 'skills' to standardize how agents perform specific tasks. Four embedded engineers worked directly with teams to spread what worked.
The 15% coding-time figure is not a GM anomaly. A 2019 Microsoft study of 5,971 developers found coding occupied roughly 14 to 20 percent of a workday. A 2018 Stripe survey put maintenance work alone at more than 17 hours per week. Haq's point, and the reason this interview is worth reading in full, is that most productivity tools attack the smallest slice of engineering work. The full piece details how GM structured human review checkpoints, what agents actually surface from telemetry data, and why Haq said the productivity gains exceeded GM's own expectations.
[READ ORIGINAL →]