We believe monitoring and intervention to be a strong tool when dealing with complex systems. Read more here arxiv.org/abs/2502.05986
@megamor2.bsky.social @oriyoran.bsky.social

Using our method, we see consistent gains across varying difficulty levels of WhoDunitEnv. We apply our method to GovSim, a recent resource sharing environment and show an increase in both survival rates and efficiency of agents.
@giorgiopiatti

The environment features several difficulty scales and two variants of action separation, to ensure an interesting task for a wide range of agents.

To evaluate our approach we release WhoDunitEnv, a collaborative environment where agents play the role of detectives🕵️, attempting to point out a culprit out of a suspect lineup. Agents each have different pieces of information about either the suspects or the culprit.

During communication, we monitor agent uncertainty, and train a classifier that predicts task success.
If our monitor signifies a failure is likely to occur, we intervene by resetting the current communication, and allow the agents another opportunity to discuss.

Naturally, prevention is the best medicine. We introduce prevention through monitoring and intervention, inspired by similar methods used in manufacturing, cyber security, and even the human immune system
arxiv.org/abs/2502.05986

In multi-agent systems, we often rely on agents to each contribute their part to solve a task. But sometimes agents make mistakes. Those mistakes can spread through the communication channel, infecting other agents and causing a complete failure!

"One bad apple can spoil the bunch 🍎", and that's doubly true for language agents!
Our new paper shows how monitoring and intervention can prevent agents from going rogue, boosting performance by up to 20%. We're also releasing a new multi-agent environment 🕵️‍♂️