OAuth won't work for autonomous AI agents. Here's what will.

The core problem. OAuth and coarse-grained API scopes were built for deterministic software and human-driven workflows, not for autonomous agents that reason, retry, and improvise. A single agent calling 200 Resy API requests per hour until it gets banned is the new normal. The panel argues that per-call runtime authorization, informed by the agent's intent and mission context, is the only path to safe autonomy.

Why intent matters. Intent-based authorization adds a "memo field" to every agent action. Instead of a static scope like "can read Salesforce," the system evaluates whether this specific tool call aligns with the user's stated goal. Two ChatGPT sessions side by side — one shopping, one debugging production — must have radically different permissions, and only runtime context can distinguish them.

The architectural pattern. The panel converges on a "brains and hands" separation: the model (brain) operates in a sandbox with no direct API access. An egress proxy outside the sandbox enforces authorization, applies budgets, and maintains a tamper-resistant audit log. This decoupling allows deterministic policy layers to constrain non-deterministic agents.

Real failure modes. The Resy incident is the canonical example. An eager agent made 200 API calls per hour, exhausting a downstream resource without malice. The panel proposes budgets — time-limited tokens with resource caps — as a structured replacement for blunt rate limiting. Budgets give agents continuous feedback so they self-regulate, much like a person managing a scarce resource.

The organizational shift. Enterprise adopters are treating agents like new hires: start with minimal access, escalate based on trust. This requires security, engineering, and product teams to share a common vocabulary around intent, mission, and policy — a cultural change as significant as the technical one. The panel sees this shared language as a prerequisite for agentic commerce and multi-cloud interoperability.

What's missing. The panel acknowledges these patterns are brand new. Most enterprises lack the mission tracking, policy infrastructure, and audit tooling required. The solutions discussed — AOTH, Keycard, agent baseline architectures — are vendor-led and largely unproven at scale. The discussion remains theoretical: no production benchmarks, no adoption numbers, only a compelling logical framework.