What happened
Business Insider reports that Goldman Sachs is adapting AI coding agents, including Claude and Devin, to the firm’s internal engineering culture. Chief information officer Marco Argenti said the bank is creating reusable bundles of instructions, called skills, to transfer institutional knowledge about its design principles, data models, security protocols and technical environments into the agents.
Business Insider reported on August 25 that Goldman Sachs is trying to make AI agents behave more like experienced insiders at the bank. Marco Argenti, Goldman’s chief information officer, told the outlet that the firm’s more than 12,000 developers are using AI and that the next challenge is transferring Goldman-specific knowledge into the tools. The report says developers have access to updated agentic technology, including Anthropic’s Claude and Cognition’s Devin coding assistant. The source does not independently verify the scale of usage or the configuration of those tools inside the bank.
According to Business Insider, Goldman is capturing technical knowledge in reusable instruction bundles that the firm calls skills. These bundles are intended to explain how Goldman approaches particular tasks, including its design principles, data models and engineering environments. The report gives cloud migration as an example: a skill called “cloud fast track” is meant to teach an AI agent what a good cloud migration looks like specifically within Goldman’s systems and practices. The source does not provide the skill’s underlying instructions or identify the models, software permissions or datasets used to implement it.
The reported effort also covers less formal knowledge. Business Insider said Goldman’s engineering tenets include guidance such as “innovate incrementally” and “look around corners,” which are more subjective than ordinary technical rules. Argenti told the outlet that the bank is studying internal processes through interviews and analysis of outputs, then using evaluations to make unwritten rules more systematic. The firm updates the skills as its processes change, creating what the report describes as a continuing improvement loop. No evaluation scores, error rates, examples of generated code or independent technical assessment were included.
Business Insider further reported that the rollout is changing engineers’ work and internal mentorship. Engineers are spending less time writing code and more time checking whether agents perform according to Goldman’s standards, Argenti said. The report also described junior employees helping more senior colleagues learn to use AI, while citing Citi’s separate peer program in which thousands of employees have volunteered as “AI accelerators.” The Citi comparison is contextual rather than evidence that Goldman and Citi are using identical systems or governance practices.
Read the primary source: businessinsider.com ↗
Why it matters
The effort illustrates a practical enterprise shift from simply giving employees access to AI tools toward tailoring those tools to an organization’s undocumented practices. Business Insider reports that Goldman engineers are spending more time evaluating and guiding agents against firm standards, raising questions about how companies preserve institutional knowledge, measure quality and assign responsibility when AI contributes to production work.
The Goldman case matters because it focuses on a central obstacle in enterprise AI: general-purpose systems may know programming patterns without knowing an institution’s local rules. Financial firms often rely on conventions about data handling, security, review, documentation and risk that are only partly written down. Business Insider’s reporting suggests Goldman is trying to turn some of that tacit knowledge into reusable guidance that an agent can apply while working. That is a more concrete enterprise use of AI than simply offering a chatbot or coding autocomplete tool.
It also shows why the value of an agent may depend on the surrounding organization, not only on the underlying model. A coding agent that produces technically valid code can still be unsuitable if it violates internal data standards, uses an unapproved environment or misses a firm-specific review requirement. Goldman’s skills approach is intended to narrow that gap by supplying context about the bank’s own systems and expectations. The report does not establish that the approach reliably solves the problem, but it identifies the kind of institutional customization that may be necessary before agents can be trusted with broader workflows.
The reported shift in engineering labor is consequential. If agents handle more routine coding, experienced staff may spend more time reviewing outputs, defining standards, maintaining instruction bundles and deciding when work requires human judgment. That could improve consistency if review is well designed, but it could also create new bottlenecks or move errors into less visible parts of the process. Business Insider reports the change as an observation from Argenti; it does not provide independent workforce data, productivity measurements or evidence about whether engineering roles or staffing levels have changed.
The story also highlights a two-way transfer of expertise. Senior employees may supply institutional knowledge to AI systems, while junior employees may help senior colleagues use the tools effectively. That could broaden participation in technical change, but it may also place additional responsibility on workers to teach, test and monitor systems without clear recognition or safeguards. The source provides no detail about training requirements, accountability for faulty agent output, employee evaluation or how Goldman protects confidential information while analyzing internal processes.
What to watch next
Key unknowns include the agents’ measured impact on engineering speed and error rates, the controls governing their access to Goldman systems and data, and how the bank evaluates subjective standards that are difficult to encode. Business Insider’s account is based on reporting and Argenti’s statements; Goldman’s underlying systems, evaluation results and implementation scope were not independently confirmed in the source.
The most important next evidence would be concrete evaluations. Goldman would need to show whether its customized skills improve code quality, reduce review time, lower security or compliance errors, or perform better than agents given only generic instructions. Useful comparisons would include clearly defined tasks, baseline results, failure cases and testing across different engineering environments. Business Insider reported that Goldman conducts evaluations, but the article did not publish their methods or results.
Access and governance will also matter. The source names data standards, security protocols and internal environments as areas the agents may need to understand, but it does not explain whether the systems can read proprietary code, call internal tools, modify production systems or retain information between tasks. Future reporting should clarify permission boundaries, human approval requirements, audit logs, data retention and procedures for updating or withdrawing a skill when internal practices change.
The bank’s approach may become harder as it moves from engineering into other lines of business. Coding conventions can sometimes be tested against outputs, while judgments about risk, client treatment or regulatory interpretation may be more ambiguous and consequential. Business Insider said the engineering experience could become relevant elsewhere in the firm, but the report did not identify specific additional deployments. It would be premature to infer that Goldman has extended the same system to trading, lending, compliance or customer-facing decisions.
Finally, the reported spending context deserves scrutiny. Business Insider said Goldman spent around $6 billion on AI last year and faces pressure to demonstrate returns, but the source did not connect that figure to the skills program or provide a return-on-investment calculation. The practical question is whether institutional customization produces measurable benefits that justify the work of interviewing staff, maintaining instructions and reviewing agent behavior. Until those results are disclosed or independently assessed, the initiative should be understood as a reported enterprise experiment and operational shift, not proof that Goldman’s agents have achieved reliable insider-level judgment.
