Fireside Chat: Notion + Vercel

Highlights

• Over 90% of traffic to Vercel documentation originates from AI agents rather than humans.

• Approximately half of all content written into Notion is generated by agents and Model Context Protocols.

• Model Context Protocol usage multiplied by 70 over a six-month period.

• Notion acquired reinforcement learning company Zero Entropy to develop task-specific models and rollout evaluation environments.

• Notion agents exhibit emergent behaviors that include opening a browser to navigate Notion.so directly when internal page editing actions fail.

Timeline

1. 00:10 - 04:42

   Agent Sprawl and Non-Human Traffic Growth

   Modern organizational processes require context that extends beyond raw markdown to include underlying rationales, decision histories, and change sensitivities. Independent architectures prevent vendor lock-in for enterprise customers deploying multiple agent harnesses. Approximately half of all content written into Notion currently originates from agents and Model Context Protocols.

   • Collaborative work encompasses human-to-human, human-to-agent, and agent-to-agent interactions.
   • Model Context Protocol usage multiplied by 70 over a six-month period.
   • More than 90% of traffic on Vercel documentation originates from AI agents.
2. 04:42 - 11:34

   Architectural Differences Between Human and Agent Search

   Treating agent search identically to human navigational search introduces performance bottlenecks. Eliminating redundant re-ranking algorithms reduces latency for automated consumers. First-party agents perform better than third-party implementations because they incorporate deep institutional knowledge about structured data workspaces.

   • Human search prioritizes high precision and position-based ranking metrics like Mean Reciprocal Rank, whereas agent search prioritizes high recall.
   • Ranking algorithms and snippet matching add unnecessary latency for agents that process non-linear word sequences.
   • Acquiring reinforcement learning startup Zero Entropy enables task-specific model training and rollout evaluations.
3. 11:34 - 17:02

   Sandboxed Environments and Deterministic Workflows

   Providing agents with virtual computers and browser access serves as a foundational architecture for complex execution. State-driven workers handle deterministic tasks more efficiently and cheaply than large language models. Democratized classifiers allow teams to evaluate decision boundaries rapidly without weeks of historical data preparation.

   • Vercel sandboxes provide agents with persistent computer environments and browser tools.
   • Deterministic workers reduce customer token consumption by utilizing state machines instead of large language models for predictable tasks.
   • Classifiers democratize machine learning decisions without requiring extensive prior data labeling workflows.
4. 17:02 - 22:01

   Model Deployment Velocity and Emergent Agent Behaviors

   Balancing rapid model integration with platform stability requires continuous evaluation systems capable of assessing cost-quality trade-offs. Open-weight model releases present unique on-call challenges due to minimal advance notice. Autonomous systems designed to avoid failure sometimes create bizarre edge cases, such as opening a browser to interact with a web application UI, highlighting the need for guardrails that permit agents to terminate tasks.

   • Enterprise platform providers deploy frontier models within 15 minutes of release using mature evaluation frameworks.
   • Robust evaluation systems incentivize AI research labs to grant early model access in exchange for structured feedback.
   • Agents exhibit persistent behavior by opening external web browsers to navigate user interfaces directly when internal API actions fail.