A reading path

Agents in live operations

An agent in a live system needs more than a good answer. It needs a bounded job, a way to ask for help, and a cost you can understand. These notes examine that boundary from model selection through the interfaces people actually use.

Follow the thinking

4 field notes

  1. The browse surface

    Chesky tried Airbnb on the agents people use. Walmart's chat checkout converted worse. ASOS still sends shoppers to look. The product is the browse surface.

    agents

  2. SpaceX bought Cursor. The interesting part is what shipped after.

    SpaceX closed Cursor (Anysphere) Aug 14. After close: Rollouts, Security Reviewer, Self-Hosted Machines, Origin — mapped to a circular SDLC. Operate and Learn still open.

    factoriesagents

  3. Same-day Sol and Opus — Opus at AA 58 is not your default

    Opus tops Artificial Analysis at 58; Sol 48, Luna 37. Same-day dual drop. Route on your harness and $/task — not Tuesday's Index. Gateway A/B without swapping keys.

    agentshitl

  4. Pace the frontier — and the agents you already shipped

    Amodei asks frontier labs to pace capability growth. OAI-HF, cheap near-frontier models, and Anthropic's hiring mix make the operator translation local: contracts, HITL, evals, and spend/send/mutate gates.

    factorieshitl