Self-Hosting Agents: What Changes When Models Become Infrastructure

[Soypete]

Talk proposal for AGNTCon + MCPCon North America 2026

Abstract

Self-hosting agents turns models into infrastructure. This talk shares operational lessons from running open models in production agent systems.

Summary

AI memory is usually treated as more chat history, more vector search, or a bigger context window. This talk reframes personalization as governed data access and shows how reliable agents need scoped retrieval, provenance, permissions, and evaluation instead of unbounded memory injection.

Changes from Previous Version

Rewrites the talk to align with the AGNTCon CFP thesis:

• New title: 'Self-Hosting Agents: What Changes When Models Become Infrastructure'
• New 5-part structure: API Hides the System → What Self-Hosting Exposes → Why Agents Make This Harder → Lessons from Pedro → Practical Architecture
• Cut external image slide and reduced benchmark slides
• Added 'When Self-Hosting Makes Sense/Does Not Make Sense' slides
• Updated final takeaway to focus on infrastructure rather than cost savings