Stateless agents, iframe tooling, efficient training, cheaper inference, and governance

Runtime: MCP goes stateless; Baseten courts lab partners reports MCP shifting to a stateless request/response core and Baseten launching Model Labs to help closed‑weight models distribute and monetize without bespoke infra. Outcome engineers get a simpler agent transport to build on—stateless MCP reduces orchestration complexity and Baseten’s Model Labs offer a pattern for integrating proprietary models into agent pipelines (Principles 07, 11).

datasette-apps 0.2a0 adds invisible‑iframe debugging and app‑listing tools so agents can test, sandbox, and manage apps programmatically. That gives teams a practical sandbox and observability surface for agent tooling and integration testing—critical for reliable agentic workflows (Principles 03, 07).

Explorative modeling: Train on the best of K guesses introduces an exploration axis that trains on the best of K samples, boosting sample, FLOP, and parameter efficiency across images, video, and language. That technique changes how you design model training for agentic tasks—upleveling exploration in training reduces data needs and can speed iteration on agents’ internal models (Principles 12, 16).

Running Kimi K3 on MI355X at Better Performance per Dollar Than B300 demonstrates MI355X delivering superior performance‑per‑dollar for Kimi K3 inference versus B300/B200. That shifts deployment economics—cheaper frontier‑model inference lets you run more aggressive agent deployments at lower cost or move latency‑sensitive components on‑prem (Principles 02, 12).

Open letters about AI development aggregates competing industry letters about open weights, distillation, and calls to pace development. Those policy fights will affect model access, release practices, and compliance for outcome engineers—plan for changing constraints around model choice and distillation workflows (Principles 10, 15).