The development
What happened
Google and Artificial Analysis have introduced Multimodal Knowledge Agent (MKA), a benchmark for AI agents working across visually complex files. It tests projects such as interpreting drawings and producing reports or spreadsheets, rather than treating professional work as text-only question answering.
Key facts
Google says domain experts across 10 verticals and 83 occupations created 200 evaluation tasks. Each pairs multimodal inputs with a project described as taking an experienced professional three to eight hours.
Expert-written rubrics contain roughly 70 yes/no checks, weighted 75% toward factual accuracy and 25% toward visual design. Models use the same Stirrup harness, tools, sandbox and turn budget.
A non-Google-model autorater inspects deliverables, including rendered pages and spreadsheet formulas. Google's reported 96% agreement with expert grading is a validation claim by the benchmark's creators, not TideReach replication.
Artificial Analysis keeps the 200-task evaluation set private to reduce contamination. A separate set of 10 tasks is public; inspecting those examples is not equivalent to reproducing the private evaluation.
The context
Why it matters
The design makes visual interpretation and usable file output part of AI-agent evaluation. TideReach assessment: results should be read with their harness, rubrics and grading conditions; a benchmark score does not establish reliability across every occupation or measure an organization's realized productivity.