Goldman Sachs Embraces Devin as Agentic AI Coder

Goldman Sachs has begun deploying “Devin,” an autonomous coding agent developed by London-based startup Cognition, as part of its technology workforce, Chief Information Officer Marco Argenti confirmed to CNBC. The bank plans to roll out hundreds, potentially thousands, of Devin instances working alongside its approximately 12,000 human software engineers.

Argenti described Devin as a “new employee” that will augment rather than replace human programmers, forming what he called a “hybrid workforce” model. Under this arrangement, human engineers will supervise agentic AI, utilising Devin to handle routine or repetitive tasks such as updating legacy systems, freeing them to focus on more complex problems.

Introduced last year, Devin is billed as the first AI software engineer capable of full-stack development. It functions autonomously, making decisions across coding tasks using natural language prompts. Early demos impressed observers, though some testing exposed limitations. In one evaluation, Devin completed only 3 out of 20 tasks successfully, and while it outperformed typical large language model chatbots, it still required oversight for errors and oversight.

Cognition – now valued at around $4 billion and backed by investors including Joe Lonsdale and Peter Thiel – updated Devin to version 2.1 in May. The startup says the model excels in large codebases where ample context is available.

Goldman’s move builds on its firm‑wide deployment of GS AI Assistant, a generative AI platform already used by tens of thousands of employees for drafting content, summarising documents, and analysing data. GS AI Assistant integrates multiple large language models, including OpenAI and Google Gemini.

The introduction of Devin raises broader questions about automation’s impact on the financial sector. Bloomberg Intelligence warned that AI adoption could put up to 200,000 banking jobs at risk globally over the next three to five years. However, Goldman maintains that AI agents like Devin will bolster human productivity rather than replace it. Argenti emphasised that engineers will need to “describe problems coherently, turn them into prompts and supervise”, highlighting a shift in human roles toward management and oversight of AI systems.

Within Goldman, internal data suggests productivity gains of 20 per cent among teams using AI tools, according to statements by President John Waldron. Competitors are also deepening their AI investments. JPMorgan has rolled out internal generative-AI tools to its 60,000 staff, while Morgan Stanley and Citigroup have launched assistants for client reports and document handling.

Nevertheless, specialists caution that agentic AI systems like Devin remain imperfect. Code generated by AI may introduce bugs or security vulnerabilities, necessitating stringent review protocols. Moreover, studies hint that less experienced developers might slow down when overly reliant on AI tools, prompting concerns about atrophy of critical thinking.

To address these challenges, experts advise that organisations establish robust supervision frameworks and ensure AI collaboration enhances human judgment. Argenti advocates a philosophy-infused approach to software engineering, urging developers to “study philosophy alongside technical skills” to enhance critical thinking and ethical oversight.

Goldman’s integration of Devin represents one of the most significant adoptions of agentic AI in the corporate world. By bringing Devin into its engineering ranks, the bank hopes to accelerate software development, reduce repetitive workloads, and better allocate technical talent—though the long-term effects on workforce composition and skills development remain uncertain.

Analysts emphasise that AI agents won’t eliminate human roles overnight. Jobs requiring nuanced judgment and relationship-building—such as senior developers, client-facing roles and compliance officers—are expected to remain. Yet, routine positions involving standardised tasks may face greater pressure.