AI summary by TwinMind
Ambient Agents and the New Agent Inbox ft. Harrison Chase
Sequoia Capital
8m 4s
Introduction & Speaker Background
- Harrison Chase, co-founder CEO of LangChain
- LangChain builds developer tools for agent development
- Focus on "ambient agents" - new concept beyond traditional chat agents
Current Agent Landscape vs Ambient Agents
- Most existing agents are "chat agents" - interact through chat interface, send messages, get responses
- Ambient agents definition: listen to event streams and act accordingly, potentially on multiple events simultaneously
Key Differences Between Chat and Ambient Agents
- Triggers: No longer human-initiated messages, but background events
- Scalability: Chat agents typically one-at-a-time interaction vs ambient agents handling however many events occur
- Latency requirements: Chat requires quick response or users leave, ambient agents can run much longer without immediate response
- UX considerations: Chat uses familiar chatbot interface, ambient agents need new interaction patterns since they run without user awareness
Concrete Example
- Email agent that listens to incoming emails
- Responds automatically
- Schedules meetings
- Pings team members or users
Benefits of Ambient Agents
- Scale impact: Thousands can run in background vs one-to-one chat interactions
- Complex operations: No strict latency requirements allow:
- Calling multiple tools
- Adding explicit planning steps
- Adding reflection steps
- Building more complex agent behaviors
Human-in-the-Loop Importance
- Ambient ≠ fully autonomous
- Four key interaction patterns:
- Approval/rejection: Human approves actions (e.g., customer refunds)
- Action editing: Modify suggested actions rather than simple approve/reject
- Question answering: Help agents when stuck, like assisting coworkers
- Time travel: Go back to specific steps (e.g., 10th of 100 steps), modify, give feedback
Benefits of Human-in-the-Loop
- Better results: Example - deep research agents asking clarifying questions upfront
- Trust building: Essential for actions like payments or approvals
- Memory enhancement: User interactions provide learning data for future improvement
Agent Inbox UX Prototype
- LangChain built prototype interface
- Shows when agents require actions
- Provides descriptions and detailed views
- Enables various interaction patterns
LangChain Technical Infrastructure
- LangGraph: Agent orchestration framework optimized for ambient agents
- Strong persistence layer
- Enables human interaction patterns
- Can stop/resume at any point
- Preserves entire state and previous states
- Supports waiting periods (seconds, hours, days)
- LangGraph Platform: Infrastructure for running agents
- Handles long-running nature
- Manages bursty event-driven scaling
- Accounts for flaky nature and human correction needs
- LangSmith: Observability tool for complex, long-running agents
Real-World Implementation
- Harrison's personal email agent (past year)
- Drafts responses and sends calendar invites
- Uses agent inbox interface
- Human-in-the-loop maintained
- Open source on GitHub
