Architecture AI Updates: September 22, 2026

1. Production Agents Fail on State Ownership, Not Hallucination

Vinoth Govindarajan of OpenAI argues in an InfoQ presentation that production agent failures mostly trace to harness design rather than model hallucination, and walks through the principles of a reliable harness: explicit state owners, invariants that hold across the run, control planes, and approval boundaries. He uses real deployments including OpenClaw as the case material. The framing puts the reliability burden on the surrounding system rather than on prompt quality, which is where most teams are still spending their effort. Source

2. DPACT Proposes Delegation, Policy, Auditability, and Co-Trust as the Agent Identity Primitives

Sahil Agarwal introduces the DPACT framework on the InfoQ podcast, covering identity, authorization, and security for AI agents acting on behalf of users. The framework separates delegation from authorization and treats auditability as a first-class requirement rather than a logging afterthought. Agent identity is the gap that shows up the moment an agent needs to act across systems that each authenticate a human, and DPACT is an attempt to name the pieces before the pattern calcifies. Source