AI Agents Seattle October

In October, we are welcoming Fred Patton, Senior Developer Advocate at Auth0 (Okta) | AI Governance & Ambient Agents.

Beyond Permissions: Adaptive Norms for Autonomous Agents

We keep trying to govern autonomous agents with permissions, but humans don't simply work from permissions. They work within norms.

TL;DR

The talk will feature Jev, it'll be a governable AI use case where a rich audit log establishes norms, such that when an agent deviates from that, a human must approve that deviation along with additional context which help determines whether that approval is a exception / one-off, or whether it should be added to the norms so that in future cases, no approval is needed.

Full description:

The Living Norms demo demonstrates a different model for agent autonomy: agents act independently inside established norms and ask for human authorization only when they encounter a meaningful deviation. Approved exceptions can become bounded precedent, allowing autonomy to expand over time without weakening hard safety boundaries.

Most agent deployments tend toward one of two extremes. At one end, a human approves every action. That limits risk, but it also creates a bottleneck, encourages approval fatigue, and eventually turns careful review into routine rubber-stamping. At the other end, the agent receives standing permissions and is allowed to operate on its own. That removes the bottleneck, but gives the agent broad authority regardless of whether a particular action is routine, unusual, or potentially dangerous.

Living Norms explores a third approach. Permission is not treated as the same thing as autonomy. An agent may have the ability to perform an action without automatically having the authority to perform it unattended. Inside established norms, the agent can proceed. When it moves outside those norms, it asks a human for narrowly scoped authorization for that specific action.

The important part is what happens after approval. A human is not simply saying, “yes, do this once.” The approval can also establish a bounded precedent: actions like this may be considered normal for a defined period and under specific conditions. Future decisions can use that precedent as evidence, while hard constraints remain fixed. The system becomes more autonomous through explicit human decisions rather than through broader standing permissions.

More about Fred (by Fred 🙂)

I design systems at the intersection of identity, governance, and AI—helping teams build ambient agents (continuous, event-driven AI processes) that are efficient, accountable, and policy-aware. My work centers on governable AI: ensuring systems stay powerful and auditable at scale. I'm the author of Building Ambient AI Agents: Create autonomous always-on product experiences that move beyond chat, which guides developers through building proactive, always-on AI systems.

Previously at Nstream, I built real-time streaming architectures across data, security, and health. I write and speak on context-aware agents, distributed systems, and functional design—with expertise in Rust, Scala, Node.js, Python, FastAPI, event sourcing, and DDD.