Universal agent frameworks at enterprise scale

Many teams experiment with AI agents as isolated features. Few design them as reusable platform primitives.

In this talk, the speaker will share lessons from designing a universal agent framework intended to support tens of thousands of agent deployments across product lines.

Topics include:
• Designing composable "skills" as first-class primitives
• Multi-tenant context containers and memory isolation
• Model routing across LLM providers
• Trust and safety enforcement in shared infrastructure
• Versioning and upgrade strategy across live agents
• Observability patterns for LLM reasoning systems

The talk will also explore how to avoid architectural drift when multiple teams build on top of the same agent layer.

This is a platform engineering talk about treating agents as infrastructure, not features.