HaMo combines experience cards with continual learning for web agents
Original title: Harness-Model Collaboration for Long-Horizon Continual Learning
The framework beats harness-only and model-only adaptation on within-domain and cross-domain workflow shifts.
HaMo is a harness-model framework for long-horizon web agents that repeatedly run related multi-step workflows. It distills trajectories into strategy-level experience cards for immediate retrieval, while an experience-conditioned objective trains the model to interpret and execute those strategies as contexts change. A retrieval-coverage selector picks compact, complementary trajectories for parameter updates. Experiments report better retention and adaptation than harness-only or model-only methods under domain shifts.
- Experience cards provide fast non-parametric reuse of procedural strategies.
- Slow parametric updates internalize how to ground and execute retrieved strategies.
- Selection favors compact, complementary trajectories for parameter updates.
- Gains are reported for both within-domain and cross-domain shifts.
From the original article
1 min read at huggingface.co
Long-horizon agents continuously encounter reusable procedural knowledge, yet existing systems typically rely either on model adaptation or external memory, trading persistent learning for rapid reuse. This limitation is especially important for Web agents that repeatedly execute related multi-step workflows.
By Zhen Bi, Zhizhi Peng, Yan Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.