Goldman Sachs has quietly turned one of Wall Street’s biggest technology bets into a working experiment on the future of software engineering. The bank’s rollout of Goldman Sachs Devin AI tools inside its technology division marks one of the clearest examples yet of a major financial institution handing real production work to an autonomous coding agent, not just a chat assistant that suggests snippets of code. And the early results, according to CIO Marco Argenti, suggest this is only the beginning.
Summary
Key takeaways
- Goldman Sachs deployed Devin, an agentic AI software engineer built by startup Cognition, across its technology division in 2025.
- Devin works alongside roughly 12,000 human engineers, autonomously scoping, coding, testing and debugging projects.
- CIO Marco Argenti says Devin performs three to four times better than previous AI tools the bank used.
- One organization using the technology saw vulnerability-fixing time drop from 30 minutes to 1.5 minutes per issue.
- Goldman plans to expand into Anthropic’s Claude in 2026 for trading, transactions and client onboarding.
- Analysts estimate up to 200,000 US banking jobs, many held by junior developers, could be at risk.
Goldman Sachs Deploys Agentic AI Software Engineer Devin
Goldman Sachs became the first major bank to put virtual software engineers to work on real production tasks rather than isolated pilot projects. In 2025, the bank deployed Devin, a tool built by AI startup Cognition, across its technology division, putting hundreds of AI agents to work alongside its roughly 12,000 human engineers and developers.
What set this apart from earlier AI coding tools banks had already been experimenting with was autonomy. Devin doesn’t just generate code when prompted. It can operate independently inside Goldman’s existing IT infrastructure while still working in tandem with human teams, a distinction that matters because it changes what the AI is actually responsible for.
Autonomous Capabilities of Devin
Rather than simply writing whatever instructions it’s given, Devin is designed to scope the full requirements of an engineering project on its own. It writes the code, tests it, submits the work for human review, and then makes corrections or bug fixes as needed. That end-to-end workflow is what separates an agentic system from a conventional AI coding assistant.
Initial Deployment and Integration with Human Teams
Goldman’s use of the technology began with modernizing legacy architecture, a well-defined and easily measured problem. It quickly became apparent, though, that Devin’s usefulness extended well beyond that starting point, prompting the bank to expand where and how it deployed agentic AI across its engineering operations.
Productivity Gains and Technical Impact of Agentic AI
The numbers behind Goldman’s experiment point to a meaningful jump in output rather than a marginal improvement. Argenti told CNBC that Devin is “like a new employee” and was expected to be three to four times more productive than the AI tools Goldman had used previously.
Reported 3-4 Times Productivity Compared to Previous AI Tools
Goldman hasn’t published detailed statistics specific to its own deployment, but a performance review Cognition carried out across its customer base in late 2025 offers a window into what’s possible. One large organization reportedly saved five to 10 percent of development time by using Devin to fix security issues in code.
Reduction in Vulnerability Fixing Time
For another customer, Cognition’s review found that the time needed to fix a vulnerability dropped from 30 minutes per issue to just 1.5 minutes when compared with human developers working alone. The system’s performance also improved with use: the share of pull requests accepted by human reviewers without significant recoding rose from about a third to roughly two-thirds over time. That trajectory matters because it suggests the technology gets more reliable the longer it’s embedded in a workflow, not just more capable at launch.
Why this matters: for a bank the size of Goldman Sachs, shaving minutes off routine vulnerability fixes across thousands of engineering tasks compounds into a substantial shift in how technology budgets and timelines get planned.
Expansion and Broader AI Integration at Goldman Sachs
Encouraged by those results, Goldman is widening its agentic AI footprint well beyond code. In 2026, the bank plans to adopt Anthropic’s Claude for work involving trades and transactions, as well as client vetting and onboarding, extending agentic automation into functions that sit much closer to the bank’s core financial operations.
Adoption of Anthropic’s Claude for Trade, Transactions, and Onboarding
Argenti told Fortune that he has shifted how he measures the technology’s impact. Instead of tracking how many staff are using AI tools, he now watches how quickly ideas turn into prototypes and then into working production models, describing the shift as the bank becoming capable of “3D printing software.” That reframing of Goldman’s Devin AI and Claude rollout, from adoption metrics to speed of delivery, signals how central agentic systems have become to the bank’s engineering strategy.
Workforce Implications and Challenges in Talent Development
The clearest tension in Goldman’s AI push isn’t technical, it’s human. Argenti has consistently framed agentic automation as a tool that frees people for higher-value work rather than replacing them outright. But that framing hasn’t settled the debate.
CIO Perspective on AI as Human Augmentation
Argenti’s position is that automation shifts human attention toward more valuable problems instead of eliminating the need for people entirely. It’s a view echoed across much of the banking sector as firms try to justify large AI investments without alarming their own workforces.
Job Loss Estimates and Risks to Junior Developer Pipeline
Researchers, including analysts at Bloomberg Intelligence, see it differently. They estimate that up to 200,000 jobs, including many held by junior-level developers, could be lost across the US banking sector as agentic AI takes on tasks that once trained early-career engineers. That estimate raises a pointed question: if junior coding roles disappear, where does the next generation of senior developers, the ones who’ll eventually lead major projects and deployments across financial services, come from?
Goldman CEO David Solomon has already spoken about restricting headcount, and Argenti himself expects AI-driven job reductions to extend beyond IT engineering into other departments. Neither executive has offered a clear answer to the pipeline question yet.
Cognition’s own review of customer deployments found a revealing limitation: Devin performs at a senior developer’s level when it comes to understanding code, but behaves more like a junior employee when it comes to execution, particularly on tasks with ambiguous or loosely defined objectives. That gap reinforces why well-scoped, structured instructions remain essential when deploying agentic systems, and it’s a caution flag for any organization tempted to hand over vaguely defined work.
Why this matters: the same efficiency gains that make agentic AI attractive to banks are the ones most likely to hollow out the entry-level roles that have traditionally trained future senior engineers, a tension that extends well beyond Goldman Sachs and into any industry scaling this kind of automation.
FAQ
What is Devin and how does it work at Goldman Sachs?
Devin is an agentic AI software engineer deployed in 2025 by Goldman Sachs, capable of autonomously scoping, coding, testing, and debugging software projects within the bank’s IT infrastructure.
How much productivity improvement has Devin achieved?
Goldman Sachs’ CIO reported that Devin is expected to be three to four times more productive than previous AI tools used by the bank.
What impact might agentic AI have on banking jobs?
Up to 200,000 banking jobs in the US, especially for junior developers, may be lost due to AI automation, posing risks for the future workforce pipeline.
How does Goldman Sachs view the role of AI in the workplace?
The bank’s CIO views agentic AI as a tool for augmenting human work, freeing employees for higher value tasks, although job displacement concerns remain significant.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
