AI BEACON #21 - Work Gets Airlocks

GitHub did not just ship a better assistant this week. It shipped a better passage into work.

Scheduled tasks, memory policy, managed plugins, sandboxes, and task APIs turn Copilot into governed execution rather than ambient help. Harvey makes the same move from the legal side, packaging provenance, ISO 42001 controls, audit logs, and human authorisation into evidence buyers can inspect. 

The workbench is becoming the moat, not the model. NHS England and BCG add the economic constraint: saved minutes do not matter if review burden and workflow redesign rise with them. 

Europe widens the frame by tying cloud, chips, AI, open source, and energy into one sovereignty package. If you have followed the Beacon, this is the point where fleets, audits, and power constraints finally enter the workflow itself. 

AI Beacon #04 first flagged agent fleets as the deployment unit; this week that pattern moved into work infrastructure. The lesson is simple: autonomy is no longer judged by model fluency alone. It is judged by whether the airlock can show owner, scope, record, and reversal.

TL;DR

·Execution rights are product design: task APIs, plugins, memory policy, and sandboxes now define what an agent may do.

Proof is becoming product: provenance, trace hooks, and human authorisation are turning trust into inspectable evidence.

Capacity is joining governance: sovereignty rules and power constraints are moving into deployment planning earlier.

Subscribe for one calm, AI Beacon each week.