Software's rethinking who counts as "the user." Connecting your favorite agent straight to your daily tools, no human clicking through every step, is going from edge case to expectation.
On a 2-3 person client project, copying a 200-person company's process sinks it before it starts.
Sprints, retros, planning ceremonies exist to coordinate people who don't talk daily. We do. Skip the theater.
Every project I ship runs auth through Better Auth now. Freelancing across different clients, I don't have time to rebuild auth from scratch each time.
This closes a gap I think about a lot: knowing how to do something versus having it actually run without you. Recording replaces documentation. That's the real shift.
Claude Cowork just added 'record a skill.' You do the task once, screen and voice recorded, and it turns into something Claude can run again. No code, no skill file to write by hand.
Ran it on a real coding task against Claude Code. Didn't expect the output to hold up this well. Still team Claude for now, but this is the first Chinese open-weight model that's made me actually pay attention.
Kimi K3 dropped this week. 2.8T parameters, largest open-weight model ever released, and it's beating Claude Opus 4.8 on coding benchmarks. Open weights land July 27.
Lo que cambia es el momento: MCP para explorar conversando con un agente. neon.ts para que la infra viva versionada y se repita igual en cada pipeline.
¿Ya conectaste un agente a tu base de datos?
La que no uso: neon.ts. Mismo principio para pipelines: proyectos, ramas y compute desde TypeScript, versionado con el código. Pensado para CI, branch-per-PR.