Notion spent over three years iterating on Custom Agents—rebuilding them four or five times—due to early limitations like weak models, short context windows, and unreliable tool-calling. A major inflection came with stronger reasoning models (e.g., Sonnet 3.6/3.7), enabling robust background execution, fine-grained permissions, and Slack-integrated workflows. Their approach centers on the 'Agent Lab' thesis: not just wrapping models, but deeply understanding collaboration patterns to build product systems around frontier capabilities. Engineering leadership emphasizes user journeys over novelty, low-ego teams comfortable deleting code, and 'demos over memos'—shipping internally first for rapid feedback. Notion organizes AI across core infrastructure, packaging teams, and a company-wide mandate that every product surface must serve both humans and agents. Their eval philosophy includes 'Frontier/Headroom' tests (designed to fail ~70% of the time) and a dedicated Model Behavior Engineer role focused on capability analysis—not just engineering. Architecturally, they evolved from XML and JavaScript agents to Markdown- and SQL-like abstractions, prioritizing progressive tool disclosure and deterministic CLI execution over opaque MCP where possible. Pricing uses credits to abstract across tokens, models, web search, and sandboxing—guided by an 'Auto' system that matches models to tasks. Crucially, Notion sees itself not as a hardware or foundation model builder, but as the system of record where collaboration data lives—powering agents, search, and workflows like Meeting Notes, which has become a key growth loop by turning conversations into structured, actionable knowledge.