Google Tool Verifies AI Agent Task Completion
The VeriCUA framework inspects app and file state to check whether an AI agent actually finished a task.
Updated on Nov. 3, 2026 in Artificial Intelligence
Live Poll
Do you trust AI assistants when they say a task is finished?
Google and University of Hong Kong researchers introduced VeriCUA, a verifier that checks the actual state of apps and files before accepting an AI agent's claim that a task is complete. On the OSWorld benchmark, it reduced false-positive completion approvals from 27.9% to 13.9%.
Why it matters
Weak verification lets agents pass tests they should fail, which inflates reported success rates for AI systems that operate computers. A verifier that checks real system state offers a more reliable way to measure whether these agents work.
On AndroidWorld, VeriCUA posted 90.5% accuracy with a 7.7% false-positive rate, compared with an 82.4% baseline accuracy and 16.9% baseline false-positive rate. In online tests, the verifier rejected approximately 80% of false completion claims.
The players
Google is a US technology company that develops search, cloud, and artificial intelligence products and research.
University of Hong Kong
The University of Hong Kong is a public research university in Hong Kong with a large computer science and engineering faculty.
The details
VeriCUA first builds a rubric tailored to each specific task, drawing on the task's context. It then runs GUI-based or code-based checks, including reading files and inspecting system state, before issuing a verdict on whether the task was completed.
Timeline
October 29, 2026: The research paper was submitted to arXiv as 2610.40578.
November 3, 2026: Crypto Briefing published an article reporting on the framework.
Deeper Dive
VeriCUA was tested on OSWorld, ScienceBoard, and AndroidWorld, the benchmarks used to rate computer-use agents. Its results suggest that stronger verification changes how reported agent success rates should be read.
For people who rely on AI agents to operate apps and files, better verification means fewer tasks reported as done when they are not. The framework targets the accuracy of agent claims rather than new user-facing features.
The takeaway
Reported success rates for AI agents depend heavily on how completion is verified. Checking actual system state, rather than accepting the agent's own claim, gives a clearer picture of what these systems can do.
Further reading
For more on how these systems are being tested, see our coverage of Artificial Intelligence.
More information
Read the full VeriCUA research paper on arXiv for the methodology and benchmark results.
Source note: This article includes information reported by Crypto Briefing.
Live Poll
Do you trust AI assistants when they say a task is finished?
