Home / Technology / AI Agents Triple GM's Autonomous Dev Output
AI Agents Triple GM's Autonomous Dev Output
28 Jul
Summary
- Software engineers at GM's AV division code only 15% of their time.
- AI agents handle 85% of tasks like data analysis and testing.
- This approach yields triple the merged pull requests and fewer defects.
Software engineers at General Motors' autonomous vehicle division now dedicate only 15% of their work hours to coding. This significant shift is attributed to the integration of AI agents, which manage the remaining 85% of development tasks. These tasks include analyzing vehicle data, troubleshooting issues, running experiments, and testing potential solutions.
GM's strategic adoption of AI agents has led to approximately three times more merged pull requests within its autonomous vehicle engineering teams. This enhancement also contributes to quicker release cycles and a decrease in defects reaching later stages of development. The company focused on redesigning entire engineering workflows around these AI agents rather than simply providing a coding assistant.
Prior to this AI integration, developers spent a substantial portion of their time on tasks outside direct coding. Research from Microsoft and Stripe indicated that developers spent minimal time writing code, with significant hours dedicated to maintenance and debugging. GM's approach involved dividing autonomous vehicle work into loops like simulation testing, road testing, and post-deployment monitoring.
The company connected its AI agents to internal tools and vast amounts of company data through custom servers. These agents were equipped with 'skills,' or instruction documents, to perform specific tasks, such as analyzing telemetry data from vehicles on public roads. The output from these agents is designed to be human-readable, providing engineers with analyzed data, potential problem identification, and supporting historical examples.
GM's agent platform is treated as a product, with dedicated engineers supporting adoption. This has accelerated feature releases while reducing bugs and test escapes. Crucially, human oversight remains at critical control points, with engineers reviewing measurements and performance metrics before advancing work towards production. GM found its productivity gains exceeded initial expectations.
