← agentic systems
Agentic · Sequential · Cookbook 11 The Python 11_sequential_pipeline.py

Three desks, one claim

Three AI agents handle an insurance claim, one after another. One summarises the email, one decides whether it is covered, and one writes to the customer. Each agent only sees what the one before it wrote, so a detail invented at the first desk ends up in the letter. Jev checks every handoff, and the claim either moves on, goes back to the same desk once, or stops for a person.

LLM agent doing the work Jev gate: one call per handoff A person A flaw that got past the gate

This claim

Press Play.

    So far

    Sent
    0
    no person needed
    Retries
    0
    sent back once
    To a person
    0
    line stopped
    Flaws missed
    0
    reached the customer
    Gate time
    0.0s
    Jev, all gates
    LLM judge
    0s
    same gates

    Move the grounding bar. Set it too low and a made-up fact gets through. Set it too high and good work gets sent back until it ends up with a person. Every claim processed so far is re-checked at the new setting.

    The code · lines this claim ran are highlighted

    
    
    
    
    
      
    Every claim so far · at the current settings
    
      
    
        
    ClaimIntakeAssessorLetterOutcome
    
        
      
    
    
    
  
    

  
    The numbers are illustrative, not measured. Gate time assumes about 0.25s per Jev call and about 4s per
  LLM-judge call. Next: 12 · Four specialists, one verdict →