Conference Presentation, Keynote
Ambient Agents and the New Agent Inbox ft. Harrison Chase
- Ambient agents are expected to listen to event streams and execute actions on multiple events simultaneously, scaling to thousands of concurrent background tasks to handle complex operations involving extensive tool use, explicit planning, and reflection steps that exceed the capabilities of one-to-one chat agents.
- Due to their long-running nature where responses are not immediately required, these agents operate with significantly less strict latency requirements compared to chat agents and are designed to pause at any point to wait for seconds, hours, or days.
- Future user interaction patterns will include approving or rejecting specific actions, editing suggested actions via advanced options, answering clarifying questions when agents encounter difficulties, and utilizing a "time travel" feature to revert to specific steps (e.g., step 10 of 100) for modification and feedback.
- Human-in-the-loop mechanisms are predicted to enhance results for deep research tasks, build trust during explicit financial or approval actions, and inform agent memory through learning from user interactions.
- The Laingraph orchestration framework is being prioritized to support persistence layers that maintain full state history, enabling the system to correct mistakes, allow state modification, and scale to handle thousands of events simultaneously.
- Observability is being addressed through Langsmith to provide visibility into the long-running and complex behaviors of these agents, while a prototype "agent inbox" allows users to view detailed descriptions of requests requiring action.
- Specific implementations include a functional email agent capable of drafting responses or sending calendar invites based on past correspondence, with open-source versions of the email agent and agent inbox intended to demonstrate future capabilities.