Back to AI Research

AI Research

HMA alternates agent sessions to sustain machine-learning experiments

Key Takeaways

  • Two agents share a workspace while session limits and context resets help test a workflow for extended engineering tasks.
  • ## Reset the session while retaining the work
  • [Humanize MLE Agents](https://arxiv.org/abs/2610.13494), or HMA, studies extended machine-learning engineering tasks.
  • The authors observe that progress and reasoning intensity can decline during a long agent session.
  • They also find that the benefits of custom harnesses can change as underlying models improve.

Reset the session while retaining the work

Humanize MLE Agents, or HMA, studies extended machine-learning engineering tasks. The authors observe that progress and reasoning intensity can decline during a long agent session. They also find that the benefits of custom harnesses can change as underlying models improve.
HMA takes a workflow approach. It alternates two agents on a shared workspace, limits the number of experiments in each session and resets context between sessions. The files and experimental work can carry forward while a new session approaches the task with a fresh context.

Evaluate the entire experimental budget

The authors model this process as a context-reset semi-Markov decision process. Their analysis describes sufficient conditions under which alternation can exceed a single-agent reference, along with experiment-budget bounds and conditions for session caps.
On 75 MLE-bench Kaggle tasks, the paper reports a 78.2% any-medal rate, eight percentage points above its strongest open-source baseline with a quarter of that baseline's time budget. Experiments with different agent pairs also report average improvements over the constituent baselines.
These results concern the paper's benchmark protocol. An any-medal rate on those tasks does not imply that a workflow wins every competition or produces a deployable model without review.

Preserve enough evidence for the next session

Franklin analysis: the shared workspace is the continuity mechanism. A handoff should retain experiment settings, measured outcomes and the current best candidate so that resetting context does not mean losing the evidence already collected.
A useful replication would compare alternation with a single agent under equal resources, report the session cap and include failed experiments in the record. It should also distinguish benefits from fresh context, different models and additional computation. HMA offers a specific way to investigate long-session slowdown, with benchmark results that can be tested against an explicitly described workflow.

Comments