Enterprises Pull Back on Agents While AI Infrastructure Keeps Spending

Two stories dominated the AI feeds this Monday, and they point in opposite directions. On one side, Goldman Sachs is pulling Claude Code from its bankers and traders - the largest enterprise retrenchment of agentic coding tooling seen so far. On the other, Anthropic committed roughly $3.4 billion to take over OnePWR's 1.7 GW queue of Caterpillar gas generators to power its data centers. Capex is accelerating even as some of the biggest customers hit pause.

The compliance wall comes for autonomous agents

Goldman's decision is the first large-scale signal that regulated firms are hitting audit and compliance walls with autonomous coding agents on the trading floor. It is not a technology failure - it is a governance one. When an agent can edit, commit, and deploy code with minimal human review, firms operating under strict regulatory scrutiny struggle to produce the paper trail their regulators expect.

The mood fits the rest of the day's discussion. A widely commented counterpoint on the AI bubble argues that AI investment is justified by real productivity gains rather than speculation - but Goldman's retreat suggests the productivity case and the deployability case are not the same thing. A tool can genuinely make bankers faster and still be unpublishable inside a regulated workflow.

Anthropic itself had a rough news cycle on the trust front: the company notified enterprise customers that a third-party cybersecurity research firm viewed internal customer account, plan, usage, and spend data in late October. Conversation content, passwords, and payment data were not accessed - but for compliance-heavy buyers, "not accessed" is exactly the kind of sentence legal teams scrutinize.

The infrastructure build-out does not care

While enterprises rein in agents at the edge, the money keeps flowing to compute and tooling. Anthropic's $3.4B generator deal is a bet that power, not chips, is the next bottleneck. And the open-source agent stack is maturing fast:

• nvke, from NVIDIA's DSX team, is an open management plane for running GPU Kubernetes on bare metal - an official path to self-hosted GPU clusters without managed-cloud lock-in.
• vercel-labs/agent-browser, a Rust CLI for browser automation built for AI agents, is trending hard (+812 stars/week) - the agent-tooling layer is standardizing quickly.
• Meta's astryx ships a design system explicitly built to be "agent ready," consumed by agents rather than only humans.

On Product Hunt the same theme repeats: NoMac tops the charts with per-minute-billed dedicated Macs for AI agents - infrastructure that only makes sense because agent workloads are real enough to rent hardware for by the minute.

Signal to watch

Watch for a second regulated firm - a bank, insurer, or broker - publicly restricting agentic tooling within the next month. If Goldman stays a one-off, this is a procurement story; if it spreads, the enterprise agent market splits cleanly into "auditable" and everything else, and vendors will have to ship for the former.