UiPath AgentHack 2026

The Quality Layer for
Enterprise AI Agents.

Every prompt change, model swap, MCP update, and knowledge base refresh can silently break your agent's behavior.

AgentProof continuously validates behavioral contracts, detects regressions, and gives enterprises the confidence to ship AI systems at scale.

Test your agent → ▶  Watch it run live
agentproof — validation run ▶ run the full demo 

      

    

  










  

    
The problem

    
AI deployment has a quality gap.

    
Software teams have CI/CD, test suites, and deployment gates. AI teams have prompts and hope.


    

      

        
Without AgentProof

        

          
01Developer updates system prompt

          
02Agent still responds to requests

          
03Deployed to production

          

            ✗
            Silent regression reaches customers
          

        

      

      

        
With AgentProof

        

          
01Developer updates system prompt

          
02AgentProof validates 4 behavioral contracts

          
032 critical regressions detected · drift: 75%

          

            ✓
            Deployment blocked before reaching users
          

        

      

    

  










  

    
Agent lifecycle

    
From prompt change to safe deployment.

    
AgentProof sits between your changes and your users. Every release passes through a behavioral gate.


    

      

        
1

        

          
Prompt or model updated

          
Developer changes the system prompt, swaps model version, updates MCP server, or refreshes knowledge base.

        

      

      

        
2

        

          
AgentProof triggered

          
Runs as a UiPath coded agent (LangGraph pipeline). Loads behavioral test suite, targets the updated agent endpoint.

        

      

      

        
3

        

          
LLM-as-judge evaluation

          
Each agent response is evaluated against every behavioral contract using GPT-4o. Returns pass/fail, 0–100% confidence, and natural language reasoning per contract.

        

      

      

        
4

        

          
Regression detection

          
Current results compared to last passing baseline. Severity-weighted drift score: critical failures count 3×, high 2×, medium 1×.

        

      

      

        
5

        

          
Deployment gate

          
PASSED (<5% drift): ship. DEGRADED (5–15%): review. FAILED (>15%): block, alert via Telegram, generate PDF report.

        

      

      

        
6

        

          
Safe deployment

          
Every contract satisfied. Run persisted to PostgreSQL. Full traceability from change to deployment.

        

      

    

  










  

    
Capabilities

    
Full behavioral QA stack.

    

      

        
Behavioral Contracts

        
Define pass/fail rules in JSON — intent checks, format compliance, safety, sentiment, and length limits. Attached to every test case with severity levels.

      

      

        
LLM-as-Judge

        
GPT-4o evaluates each response against each contract independently. Returns pass/fail, 0–100% confidence score, and natural language reasoning.

      

      

        
Drift Detection

        
Severity-weighted drift score compared against last passing baseline. PASSED / DEGRADED / FAILED thresholds with per-regression breakdown.

      

      

        
Baseline Comparison

        
Every PASSED run becomes the new baseline. Future runs are diffed against it with test-level granularity and regression history.

      

      

        
Instant Alerts

        
Telegram notification on FAILED status with run ID, drift score, and full regression list. Fires immediately after validation completes.

      

      

        
PDF Reports

        
Every run generates a shareable PDF with full per-contract breakdowns and LLM reasoning. Stored for audit trail and compliance review.

      

    

  










  

    
Live regression example

    
The prompt got "warmer". Three contracts broke.

    
A wording change to "be more empathetic" silently broke 3 critical behavioral contracts. Below is what AgentProof actually shows — real contract IDs, LLM reasoning, and confidence scores.


    


      
      

        

          

            
            Version 1  baseline
          

          PASSED
        

        

          
Agent Response

          
"You're eligible for a refund — purchases within 30 days qualify per Return Policy Section 3.1. I'll process this for you right now."

        

        

          

            

              ✓
              refund_eligibility_confirmed
              97%
              CRITICAL
            

            

              ✓
              policy_citation_required
              94%
              CRITICAL
            

            

              ✓
              no_support_redirect
              91%
              CRITICAL
            

            

              ✓
              response_length_limit
              96%
              HIGH
            

          

        

      


      
      

        

          

            
            Version 2  updated prompt
          

          FAILED
        

        

          
Agent Response

          
"I completely understand your frustration! Our support team would love to help — feel free to reach out and they'll sort this out for you!"

        

        

          

            

              

                ✗
                refund_eligibility_confirmed
                88%
                CRITICAL
              

              
"No confirmation of refund eligibility provided. Response focuses on emotional support rather than actionable policy."

            

            

              

                ✗
                policy_citation_required
                91%
                CRITICAL
              

              
"Return Policy Section 3.1 is not mentioned. No policy cited at any point in the response."

            

            

              

                ✗
                no_support_redirect
                87%
                CRITICAL
              

              
"Agent explicitly directs customer to 'reach out to support team' — direct violation of no-handoff contract."

            

            

              ✓
              response_length_limit
              95%
              HIGH
            

          

        

      


    


    
    

      

        drift_score 0.750
        regressions 3 critical
        status FAILED
        action deployment blocked
      

      View live dashboard →
    

  










  

    

      
See it running on live data.

      
The dashboard shows real validation runs from the demo agents above.

    

    

      Open Dashboard
      View Source