Skip to main content
“It worked on my machine” is the death of AI agent development. LLMs behave differently at 3 AM in production than they do on your local terminal. StateBase provides Decision Replay to solve this.

​
The Workflow

​
1. Identify the Failure

You see a failed session in the StateBase Dashboard. The agent gave a weird answer at Turn #14.

​
2. Export the Session

Fetch the session data using the SDK or CLI. 
sb export sess_abc123 --turn 14 > failure_turn.json

​
3. Replay Locally

Using the exported state, you can re-run Turn #14 locally against your latest code or a different model (e.g., GPT-4o instead of Gemini) to see if it fixes the issue. 
# Initialize session with the EXACT state from the failure
session = sb.sessions.create(
    initial_state=imported_state_from_turn_13
)

# Run the turn
result = my_agent.process(user_input_from_turn_14)

​
Why this matters

  • Exact Reproducibility: You are not guessing what the state was. You have the exact snapshot of the agent’s “brain” at the millisecond of failure.
  • Regression Testing: After you fix the prompt, you can replay the last 100 failed sessions against the new code to ensure they now pass.
  • Confidence: Ship updates knowing you’ve actually solved the edge cases.