Partnership on AI: Agent 'Action' Tools Hit 65% as Oversight Lags

A squeeze is building inside every enterprise that quietly deployed AI agents this year, and a new op-ed in Tech Policy Press puts numbers on it. Madhulika Srikumar of the Partnership on AI and Vinh X. Nguyen of the Council on Foreign Relations report that the share of "action" tools an agent uses to directly edit files, run code, or execute financial transactions rose from 27% to 65% between November 2024 and February 2026. The monitoring layer meant to sit on top of that has not scaled at anything like the same pace.

The regulatory backdrop is the EU AI Act, which requires that high-risk AI systems be "effectively overseen" by humans, and the authors argue that real-time oversight is expensive in a very literal sense. Their examples of what happens when the oversight lags are specific. An internal Meta agent exposed sensitive user and company data for two hours before it was caught. A zero-click vulnerability in Microsoft 365 Copilot let an attacker instruct the agent by email to exfiltrate a user's sensitive files. Real-time "constitutional classifiers," the filters that catch policy violations in model output, can add over 20% to inference costs. Anthropic has reportedly pulled that overhead down from 23% to about 1% by tiering the work, reserving the expensive analysis for the small share that gets flagged as risky, a pattern the piece compares to network monitoring where you log 1 in 1000 packets rather than all of them.

The other pressure point is human. The authors cite 71% of Security Operations Center staff experiencing burnout, which is roughly what unfiltered logging of every reasoning step and tool call would look like at agent scale.

The honest caveat is that this is an argument, not a survey. Its central prescription, that enterprises tier monitoring by deployment risk and capture meaningful logs at the right level of detail, leans on a small number of vendor-friendly examples rather than a broad benchmark. It does not tell you which regulators will accept which tiering scheme, or how a smaller AI shop without Anthropic-scale research could match a 1% overhead figure.

The closing line is the one worth keeping. The infrastructure to monitor these systems will get built. The only question the authors leave open is whether enterprises start now or wait for the high-profile public failure that forces them to.