eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
Goldman Sachs is testing Devin, an autonomous AI agent embedded into its 12,000-person engineering team. The tool is part of the firm’s push to create a hybrid workforce that blends human oversight with machine execution.
In comments to CNBC, Chief Information Officer Marco Argenti said the AI will take on real software development tasks, starting with low-level engineering work. While Devin doesn’t require training, vacations, or breaks, it will be closely supervised by Goldman’s developers as the bank explores the future of human-AI collaboration.
AI joins the engineering ranks at Goldman Sachs
Developed by Cognition, Devin is a software-based AI agent designed to handle full-cycle engineering tasks with minimal human input. It can complete end-to-end coding assignments, from reading documentation to writing, debugging, testing, and reporting results.
Devin is handling routine tasks at Goldman Sachs, such as updating internal codebases to newer programming languages. Assignments often considered tedious by human developers are now being offloaded to the agent as part of a larger strategy to free up engineers for higher-value work.
Goldman sees this as a lever for scale. Argenti told CNBC the firm will begin with hundreds of Devin instances and may ramp up to thousands, depending on outcomes and internal demand. He added that this agentic approach could triple or quadruple the impact of previous AI solutions, accelerating output while keeping humans in supervisory control.
The reality check: Can Devin really code?
While Goldman is optimistic, Devin enters the enterprise world with mixed early reviews.
In independent tests conducted in January and reported by The Register, the agent successfully completed just three out of 20 assigned tasks. Researchers from Answer.AI said it handled simple coding jobs but faltered on more complex ones, often pursuing infeasible solutions or hallucinating nonexistent features. They described the interface as polished, but the results as unreliable.
Cognition did not respond to The Register’s request for comment.
Goldman’s AI strategy starts with people, not products
Devin is not the firm’s first foray into artificial intelligence.
Last month, it launched GS AI Assistant, an internal tool that allows employees to interact with large language models (LLMs) behind the company firewall. Tested by over 10,000 staff, or nearly a quarter of the workforce, it was rolled out to enhance efficiency without cutting headcount.
In an op-ed first published in Fortune, Argenti made clear that tools like GS AI Assistant are just the beginning. He argued that Goldman’s next phase of AI adoption depends not just on smarter systems but on a new kind of employee: what he called “AI natives,” workers fluent in managing autonomous agents. These employees, he wrote, won’t just use AI tools but will be expected to delegate tasks, supervise results, and remain accountable for what AI delivers.