software engineers at The autonomous driving division of General Motors (GM) They spend only 15% of their time writing code, according to Rashed Haq, GM’s vice president of autonomous vehicles, in a recent interview on stage at the VB 2026 Transformation.
GM is now using AI agents to accelerate much of the other 85%, including analyzing vehicle data, classifying problems, conducting experiments and testing potential solutions.
The result, Haq said, is about three times as many pull requests merged across GM’s autonomous vehicle engineering organization, faster launches and fewer defects slipping into later stages of development.
Haq said GM achieved those advances by redesigning entire engineering workflows around agents, not simply adding an AI coding assistant.
“If you give someone just a chatbot that can code, there’s still a lot of inefficiency in that process,” Haq said during a fireside chat on the main stage of the luxurious Hotel Nia in Menlo Park, California.
Engineers spent most of their time outside of the code editor before agent AI.
GM’s 15% figure may seem surprisingly low, but research conducted before the rise of generative and agent AI came to a similar conclusion.
TO Microsoft Study 2019 Based on responses from 5,971 professional developers, they were found to spend an average of 96 minutes writing code on good workdays and 66 minutes on bad workdays. That is equivalent to approximately 20% and 14% of an eight-hour day, respectively. TO Stripe Survey 2018 found that the average developer spent more than 17 hours a week on maintenance work like debugging and refactoring.
There is no single industry benchmark: Microsoft researchers noted that previous studies placed coding between 9% and 61% of developer time, depending on how researchers defined and measured the work. But the findings reinforce Haq’s central argument. Long before agents came along, writing code represented only one part of software engineering. Accelerating that step alone leaves much of the development process intact.
“Looping it became really important,” Haq said.
GM divided its work with autonomous vehicles into several cycles: developing and testing software in simulation, testing vehicles on public roads, and monitoring the vehicles after they reach customers. Then it looked for the longest bottleneck in each loop, automated it, and repeated the process.
GM gave agents access to internal tools and data
GM connected agents to internal tools and petabytes of company data through custom Model Context Protocol (MCP) servers. It also created “skills,” or version-controlled training documents that tell agents how to perform specific tasks.
One high-value application involves telemetry collected from vehicles on public roads. Agents can analyze that data, perform an initial classification, and create issues for engineers to investigate. Through MCP connections, they can also call the underlying tools used by WebViz, GM’s system for viewing vehicle telemetry, rather than relying on the same graphical interface that a person sees.
The findings have yet to make sense to engineers. “The result has to be human readable,” Haq said. An agent could identify a potential problem, locate the affected component, search historical data for similar incidents, and provide examples to support their conclusion.
GM bases an agent’s permissions on those of the engineer who uses it.
“If an engineer is going to perform this task and needs access to these things, then their agent needs access to those things,” Haq said. “The engineer remains responsible for the agent’s performance.”
The company also uses background agents to run machine learning experiments in parallel. An engineer defines an experiment and its parameters, then agents run tests and collect the results.
Three times more pull requests, with fewer escaped defects
GM treated its internal agent platform as a product and assigned four deployed engineers to work directly with engineering teams. They helped employees identify useful workflows, spread successful practices, and adopt the tools.
Haq said the resulting increase in merged pull requests represents more than just increased code volume. “The speed at which we are releasing new features” has increased, he said, while the releases produced “fewer test leaks, bug leaks” and other problems.
People remain responsible at critical control points. Haq said GM established structured and unstructured testing and performance measurements before accelerating the broader workflow. Engineers review those measurements and determine whether each test still captures its intended goal before the work moves toward production.
Haq said GM initially expected a more modest productivity increase. “I think our only surprise was how much we could do,” he said.
GM’s approach didn’t start by giving every developer a code generator. It started by mapping the entire path from problem discovery to a verified solution in each loop (simulation, road testing, post-deployment monitoring) and then provided agents with controlled access to the tools and data needed to shorten the longest bottleneck at each stage.
