Agentic AI takes a quarter of ICML's workshop slate

ICML 2026 opened in Seoul with 23,918 paper submissions. At least 60 of the conference's 247 workshop proposals centered on agentic AI, per a July 9 recap of the conference program.

ICML workshops are where a field argues out its metrics before they harden into standards. A quarter of this year's slate is built around agents. Whatever evaluation approach survives those rooms is what builders inherit when they pick benchmarks for agent tooling over the next year.

The recap doesn't say which 60 workshops counted as agentic, or what threshold sorted a proposal into that bucket. No published cutoff. I don't know if that's a loose categorization or a tight editorial line. The gap matters, though. "Agentic AI workshop" can mean multi-agent orchestration or a narrow tool-use eval. Lumping them together hides whether the field agrees on what it's even measuring.

My take: the eval methods that come out of Seoul's agent workshops will matter more to agent frameworks than any model release this year. In my experience building with these tools, benchmarks already disagree with each other, and "success" means something different depending on who's grading. If the framework that comes out of these workshops inherits that same fuzziness, builders will spend a year optimizing agents against the wrong yardstick. They won't notice until benchmark scores and real-world reliability drift apart. Watch which workshop's framework other conferences start citing by year's end.