An agent can revise its workflow until one difficult task succeeds. That experience may still do nothing for the next unfamiliar task.
Harness², introduced October 9, combines task-level harness edits with a higher-level process that extracts experience for future adaptation. Its authors report improvements across professional benchmarks and several model-harness configurations. Read the research paper.
The workflow is another place for an agent to learn
An agent's behaviour depends on more than its language model. Tool selection, prompts, intermediate checks and the order of operations all shape the work it produces. Editing that execution harness can improve a task without touching model weights. The difficulty is deciding which edits are genuinely useful when the task is novel and no definitive reference answer exists. Searching more workflows can buy a better result, but that alone does not show a persistent ability to improve.
Harness² separates the current task's adaptation from the process used to adapt on future tasks. Its two levels of recursion are intended to make those outcomes distinguishable.
Propose, probe and compose an edit
At task level, the framework uses a contrastive Propose-Probe-Compose step to refine the harness without a ground-truth verifier. That recursion can receive more effort in parallel or sequentially. A second, harnessing-level process extracts experience from the edits and updates a continual-improvement harness. The second artifact governs future adaptation rather than simply preserving the current answer.
This division is useful when discussing recursive self-improvement. One loop may explore alternative ways to complete a single job. The other attempts to improve the strategy that chooses those alternatives. They require different evidence because success on the current job cannot by itself demonstrate transfer to the next one.
The evaluation measures bounded gains
Across three professional benchmarks and four model-harness configurations, the authors report that a single step improves GLM-5 and Gemini 3.5 Flash over their base harnesses. Maximum reported gains include 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB. They also report further benefits as the task-level recursion budget increases.
The budget qualification matters. A more thoroughly searched workflow can outperform a less thoroughly searched one even if neither system acquires a better persistent adaptation strategy. The higher-level claim therefore calls for fresh tasks under matched future search budgets, not just an improvement curve as additional effort is spent on the current task.
The editor is another model with a defined role
The main setup keeps Gemini 3.5 Flash and GLM-5 frozen as executors under existing harnesses. Claude Opus 4.8 is the default harness editor. A single improvement step is the main comparison, while an additional configuration gives task-level recursion a budget of five. The improvement library starts empty for each benchmark stream and changes after tasks finish.
These roles clarify what is autonomous and what persists. A fixed executor can still benefit from a more capable editing model around it, but that model’s calls belong in the resource account. The paper also notes that its indexed library is relatively simple, leaving consolidation, stale-experience management and retrieval as future work as the retained adaptation history grows.
Open-world feedback remains the difficult boundary
Contrastive probing supplies a way to compare edits when an exact verifier is unavailable. It cannot make open-ended quality mechanically certain. An edit may favour an output that appears persuasive, and a retained lesson may capture a benchmark-specific preference. Inspection of regressions and independent downstream assessment help establish what the harness has really learned.
The resource question is what the extra model calls purchase: a one-off answer, a reusable workflow or a better method for adapting workflows. All can be valuable. Calling them the same form of improvement makes it harder to assess their cost, persistence and likely benefit on work the system has not yet seen.
References
Questions, news tips or corrections? Contact [email protected].
