Agent Hallucination Recovery

AI agents don’t always fail loudly. Sometimes they “drift”—entering a state where they hallucinate facts or enter infinite loops. This demo shows how to use StateBase to recover instantly.

The Scenario

A customer support agent is helping a user reset their password. The agent accidentally tries to update a null field, causing the LLM to get confused and start repeating “I’m sorry, I can’t do that” in a loop.

1. Detection

In your application logic, you detect that the last 3 outputs have been nearly identical.

2. The Rollback

Instead of killing the session, you simply roll the session state back to the turn before the error occurred.

# Detect drift...
if is_hallucinating(last_turns):
    # Rollback to 1 turn ago
    sb.sessions.rollback(session_id=session.id, turn_count=1)
    
    # Retry with a 'correction' prompt
    retry_prompt = "You encountered an error. Please try a different approach."

Why this matters

Save Revenue: You don’t lose the user’s progress. They don’t have to start the 10-minute onboarding over again.
Traceability: In the Dashboard, you’ll see a “Rollback Event,” allowing your developers to analyze the exact turn that caused the drift.
User Experience: The user sees a slight delay followed by a successful answer, rather than a “System Error” message.

See it in Action

Coming soon: Interactive Dashboard Walkthrough GIF

Agent Failure Modes Debugging with Replay

Getting Started

Core Concepts

Live Demos

API Reference

Agent Patterns

Integrations

SDKs

Production Playbook

Security & Compliance

Templates & Examples

Agent Hallucination Recovery

The Scenario

1. Detection

2. The Rollback

Why this matters

See it in Action

​The Scenario

​1. Detection

​2. The Rollback

​Why this matters

​See it in Action

The Scenario

1. Detection

2. The Rollback

Why this matters

See it in Action