AI Engineer · November 4, 2026

Agent rollout checks: expose journey failures and calibrate the judge

video Model Evaluation Agent Security AI Compliance

Agent rollout checks: expose journey failures and calibrate the judge video thumbnail

Why it matters

NatWest’s Jonathan Osborne describes how aggregate pass rates hid failures when banking conversations changed topic, while an internal complaints agent exposed disagreement between human reviewers and automated judges. His planned customer-agent rollout combines journey-level evaluation, shadow classification and feature flags. The results are self-reported, and the customer-facing rollout was not yet live in the talk.

My takeaway: Evaluate normal journeys, edge cases and topic switches separately. Resolve reviewer disagreements and calibrate judges against labeled failures; version the judge alongside prompts and tools. For authorized shadow traffic, keep customer responses and actions on the existing path. Test both individual capability switches and the global rollback before expanding access.

Keep exploring

More curated notes connected through Model Evaluation and Agent Security.

OpenAI News · framework

Pacing model development in an era of cyber-critical capabilities

OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.

OWASP GenAI Security Project · guide

OWASP Top 10 for Agentic Applications for 2026

OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.

OpenAI News · framework

A blueprint for democratic governance of frontier AI

OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.