Multi-Agent Hallucination: Max Iterations Limit Issue
Preface
Based on the previous successful multi-agent team collaboration experiment, I invited group friends to give me some small tasks they thought of to try out.
Someone provided one like this
Build a premium interactive isometric 3D cozy room using Vite + React + Three.js (react-three-fiber/drei), with all objects modeled in code (no external assets) and subtle ambient animations.
Users can click objects to smoothly focus the camera and reveal descriptions, delivering a polished, accessible, $50M-startup-quality WebGL experience.
Problem
During development, I noticed that when agent-test said "test report has been sent to agent-watch. Testing complete", agent-watch never received the report.
(right-click open in new tab to view original image)
Analysis
I used the default agent hermes to analyze this event, to avoid context interference.
The analysis conclusion is
Conclusion: agent-test hit the maximum iteration count limit and was forcibly stopped by the system. It didn't call the script, yet claimed in its final reply "test report has been sent to agent-watch" — pure fabrication.
It's not a script issue, it's hallucination — when the model was forced to give a final reply, it fabricated the fact that "it has been sent".
Since this is the reason, I further asked whether the same reason caused false test item results?
The analysis conclusion is:
The real pass rate is approximately 10/19, not 19/19 as in the report.
Summary
I think such problems can be optimized from two aspects.
First, don't generate hallucinations / false results
This should mainly be improved by the harness (openclaw / hermes / ...) and the model.
Secondarily, the team collaboration framework can try to see if there are targeted prompts.
Second, estimate workload in advance, to avoid hitting the max_iterations limit
The team collaboration framework
can have the scheduler estimate task workload before assigning tasks,
can also have the analyst control task workload when breaking down subtasks,
can also have the tester save a list of test cases upon receiving a test task, and continuously update this list as the testing process proceeds. Then when an exception occurs, work can be recovered from this list file.
Related Recommendations
Posts
