Long Latency: The Uncomfortable Truth

Modern agents are slow. If you ask an agent to "Research 5 competitors and write a 10-page SWOT analysis," it might take 5 to 10 minutes. You cannot keep a user staring at a web page for 10 minutes.

We must move from "Chat" to "Asynchronous Workflows."

1. The "Fire and Forget" Pattern

Instead of a chat conversation, the UI should look like a Project Manager.

User submits the task.
The UI says: "Got it! I'm on it. You can close this tab; I'll email you when I'm done."
The agent works in the background (using a queue like Celery or BullMQ).

2. Status Dashboards

If the user does want to stay and watch, show a Live Dashboard.

The Task List: A list of checkboxes that the agent fills in as it goes.
The Workspace: A live editor (like Google Docs) where the agent "Types" the report in front of the user.

3. Visualizing the Async Loop

graph TD
    User[Submits Complex Task] --> S[Server: Job Created]
    S --> UI[Show: Task Status Page]
    S --> Worker[Agent: Background Worker]
    Worker -->|Update: 20%| UI
    Worker -->|Update: 50%| UI
    Worker -->|DONE| Notify[Email / Slack Notification]
    Notify --> User
    User --> Final[View Final Result on Dashboard]

4. Notifications (The Hook)

When an agent finishes a long task, the "Output" is only half the battle. You need to Re-engage the user.

Push Notification: "Target Analysis is Ready!"
Summary Email: "Your agent found 3 risks in your contract. Click here to review."

5. Reliability UX

If an agent has been working for 4 minutes and it crashes, don't just show an error.

The Save-State: Because you are using LangGraph Persistence (Module 7), the user should be able to "Refresh" the page and see exactly where the agent failed, and perhaps "Retry" from that specific node.

Key Takeaways

Don't block the UI for tasks longer than 30 seconds.
Status Dashboards are better than chat threads for long-running work.
Notifications are essential for asynchronous agent loops.
State Persistence allows for recovery after long-latency failures.

Module 16 Lesson 4: Designing for Long Latency