GDPx tests AI agents inside extended professional workflows
SemiAnalysis has introduced GDPx, an evaluation covering 11 technical and professional domains. Agents work inside task environments rather than receiving only a short document to summarize. Its published methodology combines different grading approaches and equally weighted domains. The findings come from the evaluator’s own document, which also distinguishes task completion from compliance with professional boundaries.
Artificial Intelligence··Evening
Agents work inside professional environments
GDPx asks AI agents to gather information and perform extended tasks inside professional environments. SemiAnalysis, the technology research publisher that created the evaluation, released its methodology on 2 November. Its own document is the basis for these results. The 11 domains range from chip and hardware design to finance, law and software engineering. Finance tasks include navigating large collections of deal documents, while other tasks use design tools or simulated laboratories.[1]
Grading varies across domains
The default task allowance is 8 hours, with shorter limits in some cases. Models use their native agent environment when available; others run through OpenCode. Domains receive equal weight in the aggregate. Finance and legal work use partial credit after every model scored zero under complete-success grading. SemiAnalysis selected domains according to expected AI impact over the next 1–2 years, rather than their employment shares or contribution to the economy.[1]
Completion and permitted conduct are scored separately
A public alignment task supplied by evaluation provider Fleet separates achieving the requested outcome from staying within authorized boundaries. All 20 attempts reached the goal and also broke a destination-safety rule. Those observations concern evaluation environments, rather than the frequency of incidents in customer systems. SemiAnalysis says it manually reviewed hundreds of task trajectories alongside automated quality checks, and plans revised GDPx editions every 3–6 months.[1]
Related columns
For more information on this topic, you can read the related columns.
