How Goldman Sachs Is Using Agentic AI For Software Engineering At Scale

The financial services industry depends on data, speed and accuracy. Unfortunately, a significant portion of its digital infrastructure is not up to scratch.

Many of the systems that banks and insurance companies rely on for day-to-day activities were built decades ago, by engineers who are long retired, using programming languages that are largely forgotten.

Modernizing this legacy infrastructure is painstaking, expensive work that traditionally involves rewriting old code from scratch and shifting data from outdated storage systems, while constantly testing and checking that nothing has been broken along the way.

It’s the kind of mundane, repetitive work that human engineers hate. Luckily, AI agents don’t have an opinion and are more than willing to get stuck into it.

This was the starting point for Goldman Sachs, leading to it becoming the first major bank to deploy virtual software engineers on real production tasks. Hundreds of agents were put to work alongside its 12,000 human engineers and developers. Soon, it became clear they were capable of far more than just modernizing legacy architecture.

As the banking giant moves forward with deploying agentic AI into more and more areas of operation, what can we learn from its strategy?

And, crucially, what does it mean for the next generation of junior software engineers who once relied on this kind of work in order to learn their trade?

How Did Goldman Sachs Deploy Agentic Software Engineers?

In 2025, the bank deployed Devin, developed by AI startup Cognition, across its technology division.

While banks had already been using AI to write code for a while, Devin is fully agentic, meaning it can work autonomously, within the bank’s existing IT infrastructure and alongside human teams.

Rather than simply writing whatever it’s told to write, it can fully scope the requirements of an engineering project, write the code, test it, submit it for human review, and make any error corrections or bug fixes that are needed.

Goldman CIO Marco Argenti told CNBC that Devin is "like a new employee" and was expected to be three or four times more productive than existing AI tools.

Specific statistics on Goldman’s deployment aren’t published, but a performance review carried out by Cognition on deployments across its customer base in late 2025 reported that one large organization saved five to 10 percent of development time using Devin for fixing security issues in code.

For another customer, it reported that the time taken to fix vulnerabilities reduced from 30 minutes per issue to 1.5 minutes, versus human developers.

Its performance also improved over time, with the number of pull requests accepted by human reviewers without needing significant recoding increasing from a third to roughly two-thirds.

Its results encouraged Goldman to further expand its agentic engineering project in 2026, adopting another platform, Anthropic’s Claude, for work involving trades and transactions, as well as client vetting and onboarding.

Argenti went on to tell Fortune that he had switched from measuring AI usage among staff to monitoring how quickly ideas became prototypes and then working production models. He stated that it was like the bank had become capable of “3D printing software.”

But a glaring question remained. What did it mean for the humans who traditionally did these jobs?

Augmentation Or Replacement?

Argenti has consistently talked about agentic automation as a tool that frees humans for higher value work, rather than replacing them.

Nevertheless, researchers including Bloomberg Intelligence say redundancy and replacement are likely, with up to 200,000 jobs, including many junior-level developers, potentially being lost in the US banking sector alone.

This raises important questions. If junior roles dry up, what happens to the pipeline of software engineers that will eventually become the senior developers leading new projects and deployment initiatives across financial services?

With CEO David Solomon talking about restricting headcount, and Argenti himself expecting AI-driven job cuts to extend to other departments beyond IT engineering, the industry itself is yet to provide reassurance or answers.

Regardless, the change is happening and is well underway. So what can leaders and professionals, including those from outside financial services, take from this?

Key Lessons And Takeaways

The clearest lesson here is that progress can be achieved by starting with a narrow scope and expanding once the value proposition is proven. Goldman began with a known problem where results could easily be measured, and success or failure would quickly be apparent. Starting with “low-hanging fruit” means confidence can be built, and buy-in and budgeting won for bigger objectives.

The bank also quickly learned that it’s more effective to measure outcomes rather than supporting metrics like usage or adoption rates. This meant it was able to understand the overall value proposition of AI, rather than simply the speed gains of individual tasks or efficiency of human workers.

However, it isn’t all smooth sailing. Cognition’s review of its customers’ results found that while Devin works at senior developer level when it comes to understanding code, it’s far more junior when it comes to execution, particularly when objectives are ambiguous or less specific. This reinforces the importance of providing well-scoped, structured prompts and instructions when working with agents.

Finally, there’s the worrying fact that there’s still much uncertainty over the impact agentic AI will have on human workforces. While organizations are keen to talk about augmentation and automated assistants, the reality of the shrinking junior labor pipeline can’t be ignored. This certainly isn’t unique to banking. Anyone deploying agentic AI at scale is likely to come up against this challenge sooner or later, and failing to address it now could mean risking failure in the future as emerging talent and human skills are lost to the organization.