BenchFlow to launch SkillsBench at ACM CAIS, presented by Google DeepMind

Why it matters

Agent demos are easy to stage, but real skill use is hard to measure. A larger, curated task set plus an SDK and Kaggle-aligned open evals could standardize how teams track agent progress and compare approaches. DeepMind’s presence increases signal and may accelerate tooling maturity around agent skills.

BenchFlow is unveiling SkillsBench at ACM CAIS in San Francisco with a launch party presented by Google DeepMind, according to the Luma event page. The Wednesday, May 27, 7:00-9:00 PM PDT gathering is framed as an afterparty to the conference’s Agent Skills'26 workshop and a chance to connect with the Google DeepMind Developer Experience Team.

Hosted by Ivan Leo, Xiangyi Li, and Nicholas Kang and run in partnership with Kernel Labs, the event highlights BenchFlow’s push to make agent evaluation more practical for builders. Attendance on the Luma page shows 198 Going, with approval required; registrants are asked to verify token ownership with a wallet before the exact venue is shared in San Francisco. A Google Maps link on the listing points to coordinates at 37.7825, -122.3925 (map).

What BenchFlow is launching

The listing describes SkillsBench as a suite of more than 100 expert curated tasks that measure how well agents use skills across diverse and complex domains. BenchFlow is also bringing an SDK to create new benchmarks and RL environments, positioning the toolkit as both a yardstick and a way for teams to extend coverage for their own use cases.

Live programming at the launch includes:

• How to create new benchmarks and RL environments using the BenchFlow SDK
• Takeaways from building Kaggle’s new Agent Benchmarks for open model evaluations

The event materials reference both “SkillsBench 1.1” in the title and “SkillsBench 1.0” in the description, an unresolved versioning detail on the Luma page. The afterparty sits alongside the ACM CAIS Agent Skills'26 workshop, which convenes researchers and practitioners focused on skills design, benchmarking, optimization, security, and ecosystem infrastructure.

Why they are gathering this crowd

Benchmarks have lagged the rapid iteration in agent tooling. By pairing a large, curated task set with an SDK and aligning with Kaggle’s open evaluations, BenchFlow is trying to give builders a way to measure skill use that travels beyond single demos. The DeepMind Developer Experience Team’s involvement signals an interest in bringing more developers into the evaluation loop, and the token-gated RSVP suggests the hosts are optimizing for a high-signal room.

For operators and researchers, the draw is straightforward: see how SkillsBench scopes and scores agent skills today, hear how Kaggle is structuring agent evaluations for the community, and decide whether to build on the SDK or propose new tasks. For investors watching agent platforms, who shows up and what gets demoed will be an early read on where the next layer of agent infrastructure might solidify.

The details

• Date and time: Wednesday, May 27, 7:00-9:00 PM PDT
• Location: San Francisco (exact venue shared after approval via Luma)
• RSVP: Approval required with wallet-based token verification on the event page
• Presented by: Google DeepMind
• In partnership with: Kernel Labs
• Aligned with: ACM CAIS Agent Skills'26

If you are building or evaluating agents and want a view of how the community is defining skills in 2026, this is the room BenchFlow is trying to curate.