By admin 
In a candid onstage interview at the recent VB Transform 2026 conference, Rashed Haq, General Motors’ (GM’s) Vice President of Autonomous Vehicles, unveiled a transformative shift in how his team operates, revealing that software engineers within GM’s cutting-edge autonomous driving division now dedicate a mere 15% of their time to the act of writing code. This striking revelation, delivered during a fireside chat at the opulent Hotel Nia in Menlo Park, California, underscores a profound evolution in software development, moving beyond the simple integration of AI coding assistants to a comprehensive redesign of entire engineering workflows around sophisticated AI agents. These agents are now diligently handling the remaining 85% of an engineer’s workload, encompassing critical tasks such as the intricate analysis of vast vehicle data, the efficient triage of complex problems, the systematic execution of experiments, and the meticulous testing of potential solutions. The tangible impact of this strategic overhaul, according to Haq, is a staggering threefold increase in merged pull requests within GM’s autonomous vehicle engineering organization, leading to accelerated release cycles and a significant reduction in defects that manage to slip through to later stages of development.
Haq emphasized that GM’s success stems not from simply providing engineers with a chatbot capable of generating code, but from a fundamental reimagining of the entire development process. "If you give somebody just a chatbot which can do coding, there’s still a lot of inefficiency built into that process," he stated, highlighting the limitations of isolated AI tools. This strategic pivot towards agentic AI, where AI systems act autonomously to perform complex tasks, addresses the long-standing challenge of developer inefficiency that predates the current generative AI boom.
The figure of 15% coding time might initially seem remarkably low, but it aligns with earlier research that meticulously examined the daily realities of software development. A comprehensive 2019 Microsoft study, which surveyed 5,971 professional developers, found that on "good" workdays, engineers spent an average of just 96 minutes writing code, translating to roughly 20% of an eight-hour workday. On "bad" days, this figure dropped to an average of 66 minutes, or approximately 14% of their time. Further underscoring the time-consuming nature of non-coding tasks, a 2018 Stripe survey indicated that the average developer dedicated over 17 hours per week to essential but time-intensive maintenance activities such as debugging and refactoring code. While there isn’t a universally agreed-upon industry benchmark for the exact percentage of time developers spend coding – with earlier studies varying widely from 9% to 61% depending on methodological definitions – these findings consistently reinforce Haq’s central argument: writing code is but one component of the broader software engineering landscape. Merely accelerating this single step leaves a substantial portion of the development lifecycle untouched and ripe for optimization.
GM’s strategy for achieving this optimization involved a systematic deconstruction of its autonomous vehicle work into distinct "loops." These loops represent the cyclical processes of developing and testing software in simulated environments, conducting real-world testing of vehicles on public roads, and diligently monitoring vehicle performance after they are deployed to customers. For each of these critical loops, GM meticulously identified the most significant bottleneck – the stage that consistently caused delays. Once identified, this bottleneck was targeted for automation, and the process was then repeated, creating a continuous cycle of improvement. "Doing it by loop became really important," Haq explained, illustrating the iterative and data-driven nature of their approach.
A cornerstone of GM’s agentic AI implementation is the provision of agents with secure and controlled access to the company’s internal tools and vast repositories of data, which amount to petabytes of information. This access is facilitated through customized Model Context Protocol (MCP) servers. Furthermore, GM developed version-controlled "skills" – essentially sophisticated instruction documents that guide the agents on how to perform specific tasks with precision and reliability.
One particularly impactful application of this agentic AI is in the analysis of telemetry data collected from autonomous vehicles operating on public roads. These AI agents are now capable of sifting through this massive stream of data, performing an initial triage to identify anomalies or potential issues, and then automatically generating detailed issue reports for human engineers to investigate further. Leveraging the MCP connections, these agents can directly interact with the underlying tools used by WebViz, GM’s proprietary system for visualizing vehicle telemetry. This bypasses the need for agents to interpret data through the same graphical user interface that human engineers use, enabling a more direct and efficient data processing pipeline.
Crucially, the output generated by these AI agents must be comprehensible and actionable for human engineers. "The output has to be human readable," Haq stressed. An agent might not only identify a potential problem but also pinpoint the affected component, meticulously search historical data for similar incidents, and provide concrete examples and supporting evidence for its conclusions, thereby expediting the diagnostic process.
GM has implemented a robust security framework by basing an agent’s permissions on those of the engineer it is assisting. "If an engineer was going to do this task and they need access to these things, then their agent needs access to those things," Haq clarified. This ensures that agents operate within the established security boundaries and do not have undue access to sensitive information or systems. He further emphasized the continued accountability of human engineers: "The engineer still is accountable for the output of the agent." This layered approach ensures both operational efficiency and rigorous oversight.
Beyond problem identification and resolution, GM also utilizes background agents to conduct machine-learning experiments in parallel. An engineer can define the parameters and objectives of an experiment, and the agents then autonomously execute the tests, collect the resulting data, and compile the outcomes. This allows for a significant acceleration of the experimentation phase, enabling faster iteration and refinement of machine learning models that are critical for autonomous driving systems.
The impact of this comprehensive agentic AI integration has been profound. GM has treated its internal agent platform as a product in itself, dedicating four full-time engineers to work directly with various engineering teams. These engineers act as liaisons, assisting employees in identifying workflows that can be optimized by agents, disseminating best practices for agent utilization, and facilitating the adoption of these new tools.
The reported threefold increase in merged pull requests is not merely a metric of higher code volume. Haq explained that it signifies a significant enhancement in the "velocity at which we’re releasing new features." This increased speed of deployment is accompanied by a reduction in what are termed "test escapes" and "bug escapes" – instances where issues are not identified during testing phases and consequently emerge in later stages of development or even in customer-facing products.
Despite the extensive automation, human oversight remains critical at key control points. GM has established rigorous structured and unstructured tests, alongside comprehensive performance measurements, to ensure the integrity of the accelerated workflows. Engineers are responsible for reviewing these measurements and verifying that each test continues to accurately capture its intended goal before any work progresses further toward production.
Initially, GM anticipated more modest gains in productivity. "I think our only surprise was how much we could do," Haq admitted, underscoring the unexpected magnitude of the improvements realized. The company’s transformative approach did not commence with simply equipping every developer with a code-generating tool. Instead, it began with a meticulous mapping of the entire journey from problem identification to the verification of a solution across each critical loop of development – simulation, road testing, and post-deployment monitoring. This foundational understanding allowed GM to strategically deploy agents with controlled access to the precise tools and data necessary to effectively shorten the most significant bottlenecks at every stage, fundamentally reshaping the future of autonomous vehicle engineering.
