From chatbot to agent

AI systems are beginning to use computers and tools to complete sequences of actions. On OSWorld, agent performance rose from roughly 12% to 66.3% in a year, yet agents still fail about one in three structured attempts.

Goal · tools · review. A simplified agent cycle with outcome review. — A simplified agent cycle with outcome review.
Museum of AI
Enlarge visual

Why it is in the museum

This room separates what is already happening, strong trends and genuinely uncertain futures.

Evidence for this exhibit

Institutional / scientific source

This source supports the dated event and the core claim described in this exhibit.

Main source: Stanford HAI, AI Index Report 2026, Technical Performance.

Open source

What to keep in mind: The evidence is used only for the claim stated on this page; broader interpretations remain curatorial.

Ask the exhibit

This small guide uses only the documented facts on this page. If your question goes beyond them, it will say so instead of filling the gap.

Start with one of the questions above.