Themes

The load-bearing ideas across both threads.

• The discovery-to-governance lifecycle — Intent → retrieve → select/bind → authorise → execute → record evidence.
• Agent-ready, not agent-trusting — Understanding a tool is never authority to use it.
• The missing collaborative-learning layer — Orchestration standards don't define how institutions jointly improve a model.
• Failure taxonomy — Seven failure classes, each with its own controls — not just 'stale data'.
• Evaluation moved from format to execution — Benchmarks shifted from argument syntax to state, policy and outcome reliability.
• Recommendations and roadmap — Harden the substrate in stages: inventory → control plane → institutional capability.
• UK public-sector implications — Roles, SLAs and liability shift as agents move from answering to acting.