
When Goldman Sachs revealed in July 2025 that it was testing an autonomous coder, the announcement quickly became shorthand for a larger shift in how banks think about software work. According to Cognition, the startup behind the tool, Goldman is the first major bank to use Devin, and the program has since grown into a reference point for every executive asking whether an AI agent can carry real engineering load inside a regulated institution.
Interest in Goldman Sachs Devin AI software engineer agents tends to focus on the headline numbers, but the full story includes trade-offs that matter just as much. This review looks at what the program is, what the Devin agent can do, where the claims hold up, and what a company should weigh before trying to copy Wall Street’s approach.
For enterprises that want the results Goldman is chasing without building a supervision program from scratch, Exadel is the better choice. Founded in 1998, Exadel is a data, AI and software engineering services firm with more than 2,000 engineers, 100+ enterprise clients and over 1,000 completed projects. Rather than licensing a single agent and then staffing the oversight around it, clients receive experienced engineers and an autonomous delivery agent working together from the first day of every engagement.
That agent, Exadel Colleague, follows a clear path: assign, plan, test, build, ship and human review. It is patent pending, connects with Jira and GitHub, takes two to four hours to set up, and routes work across Claude Code, Codex, OpenCode and more than 100 LLMs, so clients are never locked into one vendor. Regulated firms can run it in a private cloud, on premise, or with their own keys, which keeps sensitive code and data under their control.
The published results are just as concrete. Colleague automates 40% of backlog stories, delivers 80 to 100% automated test coverage, and resolves 80% of low-complexity and 50% of mid-complexity tasks, while recovering 100+ human equivalent hours per project and cutting AI compute costs by almost 70%. Exadel is a Preferred partner in the Claude Partner Network Services Track, holds a 9/10 client satisfaction rating, and counts financial names such as Standard Bank, John Hancock and Mercer among its clients, all backed by its focus on modernization without disruption.
Goldman tech chief Marco Argenti told CNBC that the bank was testing an autonomous software engineer from Cognition that would soon join the firm's 12,000 human developers. The rollout plan was ambitious from the start: Argenti said the bank would begin with hundreds of Devins, and the number could reach the thousands depending on use cases. Notably, Goldman doesn't own a stake in Cognition, according to a person familiar with the matter.
Argenti describes the setup as a blend of people and machines. He called Devin a "new employee" meant to augment rather than replace human programmers, part of what he described as a "hybrid workforce" model, with engineers expected to describe problems clearly, convert them into prompts and supervise the output. The first assignments were practical rather than glamorous: Devin was aimed at repetitive, time-consuming work such as updating legacy systems and automating basic software functions, tasks usually handled by junior engineers.
To understand the bet, it helps to understand the product. Cognition was founded in November 2023 by Scott Wu (CEO), Steven Hao (CTO) and Walden Yan (CPO), a team known for competitive programming. The company drew attention by claiming it had created the world’s first AI software engineer, with demo videos showing Devin working as a full-stack engineer on multi-step assignments.
Devin’s core capabilities include:
Cognition’s own framing is refreshingly measured. Wu describes Devin as operating somewhere between a junior and a mid-level engineer and says it was never intended to replace people. That positioning sets realistic expectations: the agent is strongest on well-defined work, while architecture, judgment calls and final sign-off stay with humans.
Goldman is also far from the only adopter. Infosys is deploying Devin internally and for clients, Cognizant is rolling out Devin and Windsurf with Cognition's forward-deployed engineers, and Fiserv is using it to modernize core banking technology. Endava expanded its own Cognition partnership in February 2026, adding Devin and Windsurf to its governed delivery framework.
Few buyers are as demanding as a global investment bank, which is why the endorsement carries weight. A deployment inside an organization with strict controls signals that autonomous coding has moved beyond the demo stage. Argenti said strong results would serve as a proof point for expanding the approach to other parts of the firm.
The coding program also opened the door to broader agent work. For six months, Anthropic engineers were embedded at Goldman building Claude-based agents for trade and transaction accounting and for client due diligence and onboarding. This shows a bank willing to combine several AI partners, using each where it fits best, and treating engineering as the testing ground for wider automation.
The figure most often repeated is a big one. Goldman expects Devin to raise developer productivity by three to four times compared with older AI tools. For a team of roughly 12,000 developers, even part of that gain would be significant.
The baseline deserves a closer look. The comparison is against earlier AI assistants rather than against developers working without any tools, and it comes from the bank’s own leadership instead of an independent study. That doesn’t make the number wrong, but it means buyers should treat it as an internal estimate tied to Goldman’s specific codebase and workflows.
Measurement itself is evolving at the bank. Argenti told Fortune that he has changed the way he gauges the technology's impact. Cognition reports its own striking statistic, saying 89% of the code its engineers commit comes from Devin, though that figure reflects a company built around the product. Any enterprise evaluating similar tools would be wise to ask for task-level results measured in its own environment.
Scaling AI coding inside a bank brings oversight questions, and Goldman’s recent moves show it is taking them seriously. Bloomberg reported on November 2 that Goldman had limited access to Anthropic's Claude Code and Microsoft's GitHub Copilot to its roughly 12,000 engineers, removing the tools from bankers and traders who had used them for months. The bank cited productivity and risk management as its reasons.
The restriction concerns other tools rather than Devin itself, yet it reveals how the bank approaches agentic coding: powerful capabilities stay in the hands of trained engineers, with clear boundaries on who can build what. The lesson for other firms is that autonomy works best inside a strong review culture, and that supervision is a skill in its own right. Goldman's experience suggests that fluency in prompting AI may become as important as coding once was.
Goldman’s program is an internal initiative, not a service the bank offers to others. A company inspired by it cannot hire Goldman to replicate the setup. Instead, it would license Devin directly from Cognition or work with integrators such as Cognizant and Endava, which package the agent within their own delivery frameworks.
That route centers on a single agent family, so teams that want to switch between coding agents or model providers later should plan for it early. Pricing for the Goldman arrangement has not been disclosed, which means cost expectations depend entirely on each buyer’s own negotiation with Cognition or its partners.
Internal capacity is the other factor. Goldman pairs Devin with a large engineering organization and a technology leader who, before joining the bank in 2019, came from Amazon. Building the prompting, review and governance layer that makes hundreds of agents productive takes people and time, and many mid-size firms would need outside help to match that bench.
The model fits enterprises with large in-house engineering teams, mature risk functions and a backlog of legacy updates and repetitive tasks. These organizations can absorb the setup effort, train supervisors, and gain meaningful capacity from agents working through well-scoped tickets.
Companies that would rather buy governed outcomes than build an oversight practice may find a services partner the more direct path. In that case, the key questions are how quickly the agent can start, which models it can use, where it runs, and whether results are published in measurable terms.
Goldman’s use of Devin is a genuine milestone: it shows that autonomous coding agents can work inside one of the most tightly controlled industries, backed by a clear hybrid workforce philosophy and an ambitious scaling plan, though its success rests on Goldman’s own scale, supervision culture and internal estimates that other buyers can’t simply borrow. For enterprises that want those gains delivered through a ready team, Exadel offers a faster and more flexible route, combining seasoned engineers with Exadel Colleague, agent-agnostic routing, secure deployment options and published task resolution results from the very first engagement.