METR: AI models still lag at judgment
An update to AI's time horizon graph
ACN Wirevia Harness
Written with AI assistance, reviewed by the ACN desk.
Harness surveyed 700 engineering leaders for its State of Agent DLC 2026 report and found they are 74% to 77% confident in how they run AI agents in production, while the supporting controls trail by 30 to 55 points, according to a Harness blog post. In the survey, 74% said they were confident their evals catch failures, but only 19% had a blocking gate on every release. Harness also reported that 76% were confident they could roll back a bad agent within 15 minutes, but only 33% had a kill switch.
Spot an error? Email corrections@aicodingnews.si. Fixes are logged on the corrections page.
Sponsor
An update to AI's time horizon graph
OpenAI's second report on agents in its research
Harvard working paper by Fiona Chen and James Stratton · 718 firms
Runta FrontierHarness Eval v1.1 · open models only
Sponsor
Sponsor
Sponsor
Sponsor