GM AI agents triple merged pull requests in autonomous vehicle engineering

General Motors has reworked parts of its autonomous vehicle engineering operation around GM AI agents, and Rashed Haq, the company’s vice president of autonomous vehicles, said the shift has produced roughly three times as many merged pull requests. Speaking at VB Transform 2026 in Menlo Park, Haq said the approach has also led to faster releases and fewer defects reaching later stages of development.

The change is aimed at the work surrounding code, rather than only code generation. Haq said software engineers in GM’s autonomous driving division spend about 15% of their time writing code, while the rest goes to tasks such as analyzing vehicle data, sorting issues, running experiments and validating fixes.

GM did not disclose the underlying pull request counts, the size of the autonomous vehicle engineering group or the cost of the internal agent platform. The metric is still notable because it ties the agent rollout to a core engineering output, rather than to a softer claim about developer satisfaction or time saved.

How is GM using AI agents?

Haq said GM divided autonomous vehicle development into several operating loops: software development and testing in simulation, public-road vehicle testing and monitoring after vehicles reach customers. Teams then looked for the slowest step in each loop, automated that bottleneck and repeated the process.

The company connected agents to internal systems and large volumes of company data through customized Model Context Protocol servers. GM also created version-controlled “skills,” described as instruction documents that tell agents how to complete particular tasks.

One use case is telemetry from vehicles operating on public roads. Haq said agents can review that data, conduct an initial triage and open issues for engineers. Through GM’s internal connections, agents can also call the tools behind WebViz, the company’s vehicle-telemetry visualization system, instead of operating only through the same graphical interface used by humans.

Haq said the output still has to be understandable to engineers. In practice, an agent might flag a possible issue, identify the relevant component, search prior data for similar events and provide examples that support its finding.

What controls did GM put around the agents?

GM bases an agent’s access on the permissions of the engineer using it, according to Haq. He said the engineer remains accountable for the agent’s output, which keeps review and responsibility inside the existing engineering chain.

The company also uses background agents to run machine-learning experiments in parallel. In that workflow, an engineer sets the experiment and parameters, while agents execute tests and collect results.

Haq said GM assigned four deployed engineers to work directly with internal teams on the agent platform. Their role was to help identify useful workflows, spread practices that worked and support adoption, treating the platform as an internal product rather than a tool dropped into the organization.

The company also put measurement and review steps in place before accelerating the broader process. Haq said GM uses structured and unstructured tests, along with performance measurements, and engineers decide whether a test still captures the intended goal before work progresses toward production.

Why did GM focus beyond coding assistants?

Haq framed the 15% coding figure as the reason a narrow coding chatbot would leave much of the engineering process untouched. Older research points in the same direction: a 2019 Microsoft study of 5,971 professional developers found coding time averaged 96 minutes on good workdays and 66 minutes on bad ones, while a 2018 Stripe survey found more than 17 hours a week went to maintenance work such as debugging and refactoring.

Those studies used different definitions and methods, and Microsoft researchers noted that prior estimates of coding time varied widely. The useful point for engineering leaders is narrower: GM’s reported gains came from redesigning the full path from problem detection to verified fix, then giving agents controlled access to the tools and data needed to shorten the slowest parts of that path.

This story draws on original reporting from VentureBeat.