Most teams still treat AI agents as prompts with tools. But as agents become persistent collaborators, we need a richer design discipline: agent modeling. What is this agent responsible for? What should it remember? What tools can it use? What boundaries should it respect? How does it hand work back to humans or other agents?

This talk introduces agent modeling as an emerging practice for designing useful, trustworthy AI teammates. We'll cover the core dimensions of an agent model — purpose, scope, memory, permissions, environment, feedback loops, and ownership — and share lessons from building Agor and using it internally at Preset. The practical side: how Agor helps make agent models real by combining knowledge bases, isolated branches, MCP tools, session genealogy, and shared team visibility, so agents become durable participants in a workflow rather than disposable chat threads.

Target audience: AI engineers, applied-AI and agent-infrastructure builders, devtools and platform teams.

Format & length: best as a 25–40 minute talk; adaptable to a keynote or hands-on workshop; can flex more conceptual or more hands-on (worktrees, MCP tools, context engineering, review loops).

Attendees leave with:
(1) the dimensions of an "agent model" — purpose, scope, memory, permissions, environment, feedback loops, ownership;
(2) how to design agents as durable teammates rather than disposable chat threads;
(3) the mechanics that make it real — knowledge bases, isolated branches, MCP tools, session genealogy.