Now

Building: EP 03 of Agents, Orchestrated. An agent that buys FIFA 2026 tickets overnight under an AP2 v0.2 mandate it cannot override, with the spending policy running as a decision table outside the model. The field guide is out; the episode and repo ship next.

Measuring: agent evals on live enterprise workloads. The four-axis method (cost, accuracy, tool reliability, temperature sensitivity) keeps producing findings that contradict the received wisdom, like temperature zero making an agent call more tools, not fewer.

Thinking about: sovereign AI in regulated industries. If your data cannot leave the country, your agents inherit that constraint, and most multi-agent architectures quietly assume it away. The privacy-boundary pattern in the A2A guide is my current answer.

Next in the series: multi-LLM routing decided by the orchestrator, agent-to-agent delegation at enterprise scale, and making the verification layer a deployable artifact.