Google Benchmark Tests AI Agents on 200 Visual Knowledge Tasks

Google and Artificial Analysis introduced a 200-task benchmark for AI agents, measuring how well they handle images, documents and spreadsheets in work designed to resemble professional assignments.

The Multimodal Knowledge Agent (MKA) benchmark draws on expertise from 10 sectors and 83 occupations, including architecture, biology, underwriting and data science. Each task is designed to require three to eight hours of work from an experienced professional. Agents receive files and a prompt, then produce a deliverable for evaluation.

A rubric of about 70 yes-or-no criteria scores each result. Factual accuracy accounts for 75% of the score, while visual design accounts for 25%. Models run through Stirrup, an open-source harness that gives each run a shared tool set, a sandbox and a fixed turn limit.

Testing found 96% agreement between human domain experts and the non-Google agentic grader. The full 200-task evaluation set is private to limit data leakage; a separate set of 10 tasks is public for experimentation.

Frontier models averaged between 30 and 130 turns per task. Gemini 4 Carbon ranked second in quality among the tested models.

Gemini 4 Carbon ranked second in quality while costing about one-sixth as much as the top-scoring model. Gemini 4 Carbon is available through the Gemini API in Google AI Studio.

Google has separately expanded Antigravity into research, writing, data analysis and document production. The MKA benchmark evaluates agents on visually complex professional tasks using files such as images, documents and spreadsheets.