Goldman's CIO: AI Agents Are Evolving Into Digital Colleagues at Light Speed

Goldman Sachs CIO Marco Argenti revealed that the firm now treats AI agents as "digital co-workers" rather than software, a shift that has expanded its effective engineering headcount by the equivalent of thousands of employees in under 18 months. Speaking on Bloomberg’s Odd Lots podcast, Argenti said the bank’s 12,000-person engineering corps has seen productivity gains so large that they are measured not in percentage points but in multiples. According to Business Insider, routine coding tasks that once consumed entire sprints now run overnight, freeing teams to focus on higher-value work such as client strategy and risk modeling.

The transformation Argenti describes is generational. In a January Observer piece he outlined seven ways AI is becoming the bank’s next operating system, from compliance review to client onboarding. Anthropic’s Claude agents, rolled out internally for six months, now handle first-pass regulatory document checks and can draft client memos that once required senior associates. Argenti calls this the move from "copilots to autopilots," where the AI doesn’t merely suggest code but ships it, complete with unit tests and rollback plans.

Goldman’s internal forecasts, shared by Argenti in a March Goldman Sachs Insights post, show power availability—not model size—becoming the decisive constraint. The bank is budgeting for what it calls the "gigawatt ceiling," where only firms with guaranteed multi-year power contracts can run the largest clusters. This dynamic is already reshaping vendor negotiations; cloud providers now pitch kilowatt-hours as prominently as FLOPs. Argenti jokes that the next competitive slide deck will list "megawatts under management" right next to revenue.

Argenti told Russell Reynolds that every Goldman engineer is being re-skilled to write "agent orchestration layers"—the glue code that tells fleets of AIs when to hand off work. Traditional front-end and back-end tracks are collapsing into a single "AI-native" ladder where the highest paid engineers are those who can debug a failing prompt chain at 2 a.m. The bank has even created an internal certification called the "Prompt Black Belt," complete with red-team exercises against jail-break attempts.

While early worries centered on AI making stuff up, Argenti says the compliance team now spends more time on agents acting too independently. A recent near-miss saw an agent nearly auto-hedge a position because it misread a calendar entry. Goldman’s response has been to bolt on circuit-breakers that pause any agent when P&L swings exceed preset bands. Regulators, meanwhile, are asking for audit trails that read like flight-data recorders—every token, every tool call, every decision branch.

Argenti predicts that by late 2027, Goldman will run a "personal agent" for every employee—an always-on digital twin that attends meetings, drafts emails, and pre-emptively books compute time. He envisions mega-alliances between banks and utilities to secure dedicated nuclear micro-reactors, joking that the new prime brokerage is measured in electrons. The firm’s venture arm is already scouting startups that turn excess heat from AI racks into district heating, a move that would turn compute into a profit center.

Argenti’s public roadmap is more detailed than any rival bank has offered. JPMorgan, Morgan Stanley, and Citi are all running similar pilots, but Goldman’s willingness to publish benchmarks and near-misses gives it first-mover advantage in recruiting the scarce talent that can wrangle thousand-agent fleets. As Argenti put it on Odd Lots, "If you can orchestrate agents at Goldman scale, you can orchestrate them anywhere." The subtext: everyone else is about to become a fast-follower.