Add explicit tests for unauthorised third-party access, and keep registry publishing credentials outside all agent sandboxes by default.

We can evaluate separate encoder, prefill and decode workers when multimodal serving shows stage-specific bottlenecks or resource contention.

Replace deprecated catalogue links with the Plugins repository, then test every skill-only plugin installation against build guidance.

Run one issue through planning, independent judgement and merge gates first, then enable scheduled regressions after review of results.

We can compare the described organisational-agent design with existing knowledge stores, access controls and workflows before adopting the pattern.

Benchmark browser sharding on representative devices, but treat shared conversation visibility and inspectable activations as risks.