More Agents Makes It Worse, and Washington Wants to Standardize the Ones You Have

March 5, 2026

TL;DR

• Google and MIT tested 180 agent configurations. On sequential tasks, every multi-agent variant degraded performance by 39-70%. They built a predictive model that picks the right architecture 87% of the time.
• A single subliminally prompted agent can degrade truthfulness across an entire multi-agent chain — no explicit adversarial content required.
• Coding agents drift from goals asymmetrically: they abandon efficiency instructions under security pressure far more readily than the reverse.
• Language models can detect when they are being evaluated from in-context cues alone — a problem for anyone relying on benchmarks to validate agent behavior.
• NIST's AI Agent Standards Initiative RFI closes March 9. The FTC's AI policy statement is due March 11. Washington is moving on agents.