SF : Dust x Anthropic: It's called having Skills

Headed to 972 Mission Street for the Dust and Anthropic event on Skills. The premise lands quickly. Instead of fine tuning a model or stuffing a system prompt with everything you might need, you give an agent modular packages of expertise it loads on demand and drops when the job is done. The DRIs of Skills at both companies walked through the architecture and the design tradeoffs, and the live demo showed an agent shifting from generalist to domain specialist in roughly three seconds, then executing a real domain workflow end to end. Anthropic released the Skills specification as an open standard around this period, and the partner directory already includes skills built by Atlassian, Figma, Canva, Stripe, Notion, and Zapier, which is a strong signal that this is turning into an interoperability layer rather than a single vendor capability. The thing I valued most was hearing the design philosophy from the people who own the call. The hard parts are not getting an agent to call a skill, they are deciding what belongs inside a skill, when to invoke one, how to compose them, and how to keep them small enough that the model can reason about which one to pick. This direction is where most production agent stacks end up, and the talk made the path more concrete. Strong evening, well organized by the Dust team.