ResearchLow importance
Harness-S: agents learn from “sibling harnesses” when the main harness fails
The paper proposes training agents with a fixed set of harness variants (the agent's execution environment), each specialized in fixing one kind of failure. Only when all attempts in the base harness fail does the system try the variants chosen by a router and use those trajectories in reinforcement learning. On AppWorld, SWE-Gym and DBBench, the trained policy substantially outperforms ordinary training when it goes back to the original harness.
Published in edition no. 17 of Noticiário do Joãozinho ·
Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese
Original source
https://arxiv.org/abs/2611.04277