Conference Presentation, Keynote
Ambient Agents and the New Agent Inbox ft. Harrison Chase
- Harrison Chase, CEO and co-founder of LangChain, defines ambient agents as systems that listen to event streams and act on them concurrently, rather than waiting for direct human chat input.
- Unlike "chat agents" which are limited to one-on-one interactions with strict latency requirements, ambient agents can run thousands of instances in the background to scale impact and execute complex, multi-step operations without time pressure.
- Ambient agents are explicitly not fully autonomous; they require human-in-the-loop interaction to ensure accuracy, build trust, and improve future performance.
- Four key interaction patterns enable human oversight:
- Approving or rejecting specific actions (e.g., granting refunds).
- Editing suggested actions before execution.
- Answering clarifying questions when an agent stalls or encounters ambiguity.
- "Time travel" to revert to, inspect, or modify previous steps within a long-running workflow.
- Human interaction is critical for enhancing agent memory, as learning from user corrections and feedback allows the system to adapt and perform better over time.
- LangChain has prototyped an Agent Inbox UI, allowing users to view agent requests, see detailed action descriptions, and approve or modify tasks directly.
- Technical infrastructure updates include:
- LangGraph: Enhanced with a robust persistence layer that saves the entire state of an agent at any point, enabling interruption, resumption, and state rollback.
- LangGraph Platform: Designed to handle the "bursty" nature of event-triggered agents and manage long-running workflows with high scalability.
- LangSmith: Providing critical observability and visibility into long-running, complex agents to track performance and debug errors.
- Chase highlights a personal use case of an open-source email agent on GitHub that drafts responses and sends calendar invites, utilizing the human-in-the-loop Agent Inbox for final approval.