Goldman Sachs Puts Agentic AI to Work on Production Code, Not Just Prototypes

Goldman Sachs is no longer testing agentic AI in a lab. According to Forbes contributor Bernard Marr, the bank has deployed hundreds of autonomous coding agents alongside its roughly 12,000 human engineers and developers, making it the first major bank to assign virtual software engineers to real production tasks. The work began with modernizing legacy systems, but it has expanded into trading, transactions and client onboarding.

What Goldman actually deployed

In 2025, Goldman’s technology division deployed Devin, the agentic coding assistant built by AI startup Cognition. Unlike earlier code-generation tools that produce snippets when prompted, Devin operates autonomously within Goldman’s existing IT environment: it scopes an engineering task, writes the code, tests it, submits it for human review, and handles error corrections and bug fixes.

The initial target was legacy modernization. Financial infrastructure often relies on code written decades ago by engineers who have retired, sometimes in programming languages few current developers know. Rewriting that code and migrating data is expensive, repetitive work. According to the Forbes report, Goldman went further once the agents proved capable.

What the early numbers show

Reported results come from Cognition’s review of customer deployments in late 2025, not from Goldman’s internal benchmarks:

| Finding | Reported detail |
| Security-fix time | One customer saw vulnerability fix time drop from about 30 minutes per issue to 1.5 minutes compared with human developers |
| Development time | One large organization saved 5-10% of development time fixing security issues with Devin |
| Code acceptance | Pull requests accepted by human reviewers without significant recoding rose from about one-third to roughly two-thirds |
| Goldman expansion | 2026 adoption of Anthropic’s Claude for trades, transactions, client vetting and onboarding |
| Open question | Goldman has not published its own performance metrics for the deployment |

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 earlier AI tools. He later told Fortune that the bank now measures how quickly ideas become prototypes and then working production models, a shift he described as “3D printing software.”

The caveat is that Cognition’s numbers come from third-party deployments, not from Goldman’s own workloads. They show potential, not verified bank results.

Why the expansion matters

The move beyond code modernization is the important signal. In 2026, Goldman added Anthropic’s Claude for business-critical areas: trades and transactions, plus client vetting and onboarding. That widens agentic AI from the technology department into operations where errors carry direct financial and compliance risk.

The bank’s stated approach is augmentation, not replacement. Argenti has consistently described agentic automation as a way to free humans for higher-value work. The measurable shift from tracking AI usage to tracking how quickly ideas reach production suggests the bank is treating the technology as a delivery mechanism, not a novelty.

The unresolved jobs question

Not everyone reads the expansion the same way. Researchers at Bloomberg Intelligence say redundancy and replacement are likely, with up to 200,000 jobs potentially lost in the US banking sector alone, including many junior-level developer roles. Goldman CEO David Solomon has talked about restricting headcount, and Argenti has said AI-driven cuts could extend beyond IT engineering.

The concern is not only current employment. Junior engineering roles have historically been the training ground for senior developers. If that pipeline shrinks, banks may face a talent problem in a generation, even as they automate more code work. No firm-wide policy or industry answer has been published.

What remains unclear

Goldman has not released its own performance data for Devin or Claude. The figures from Cognition come from its broader customer base and may not reflect the bank’s workloads, review standards or security requirements. The 200,000 job-loss estimate is a projection, not a confirmed layoff plan.

Also unresolved is how well the agents perform when goals are ambiguous. Cognition’s review found that Devin works at a senior level when understanding existing code but is far more junior at execution, especially when objectives are less specific. That suggests the quality of prompts and project scoping still determines much of the outcome.

The practical takeaway for engineering teams is to start with a narrow, measurable problem, as Goldman did, and to judge the tool by outcomes rather than usage counts. For junior developers, the labor market signal is real, but the reported numbers so far come from vendors and outside analysts, not from Goldman’s own balance sheet.

Source: https://www.forbes.com/sites/bernardmarr/2026/08/06/how-goldman-sachs-is-using-agentic-ai-for-software-engineering-at-scale/

Datos clave