$ cat wiki/papers/2026/2608.16798-clawgym-ii.md
ClawGym II: Exploring Black-Box RL on Agent Harness (arXiv:2608.16798)
TL;DR
Trains a model through a harness it does not control, by placing a serving proxy at the model boundary to capture calls, reassembling them into prefix trees and running PPO or GRPO over the recovered tree. Reports +9.98 and +14.81 Pass@1 points on ClawGym-Bench with Qwen3-30A3B through OpenClaw and Claude Code respectively, stable over 200–400 optimization steps (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED from this environment and the
paper was not read; the HuggingFace Daily Papers snapshot carries title, id, date
and abstract only
(source).
Listed on HuggingFace Daily Papers, 2026-08-19, 34 upvotes — that community's popularity signal and nothing more (source).
Method
The stated obstacle: RL through a complex harness is largely unexplored, because scaling it to long-horizon agent tasks is hard. The harness is treated as opaque — hence "black-box".
| Component | What it does |
|---|---|
| Sandbox infrastructure | isolates task environments and harnesses in temporary sandboxes for large-scale concurrent rollouts |
| Serving proxy | sits at the model boundary and captures model calls, decoupling policy optimization from opaque harness execution |
| Prefix trees | captured calls reorganised into trees to reconstruct multi-turn trajectories and improve training efficiency |
| Tree-adapted PPO / GRPO | both critic-based and critic-free objectives adapted to optimize over the recovered tree |
| Mix-harness training | a single model jointly optimized by heterogeneous harnesses |
| Training–inference consistency is stated to be maintained throughout. |
Results
All figures as reported; none independently verified (source).
| Setting | ClawGym-Bench Pass@1 gain |
|---|---|
| Qwen3-30A3B through OpenClaw | +9.98 points |
| Qwen3-30A3B through Claude Code | +14.81 points |
| Stability is stated over 200–400 optimization steps. Consistent gains are also | |
| claimed on JobBench and OfficeQA, with no figures given. |
What the abstract does not give: base Pass@1 for any configuration, the mix-harness result, wall-clock or compute cost, ClawGym-Bench's composition, or any figure for JobBench and OfficeQA.
Significance
It changes what "the harness" is in this wiki's argument. Every prior entry in Eval Harness Configuration treats the harness as a fixed scaffold wrapped around frozen weights — the whole point of StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (arXiv:2608.15089) is that the weights do not move. ClawGym II runs the causal arrow the other way: the harness becomes the environment the weights are trained against. That is a third position, and it is the one that makes the model/harness split hardest to maintain, because after training the model is no longer harness-agnostic.
The gap between the two harnesses is the number worth keeping. Same model, same benchmark, same method: +9.98 through OpenClaw, +14.81 through Claude Code — a 1.48× spread attributable to the harness alone. That is a direct measurement of the thing this cluster keeps inferring indirectly from incomparable papers, and it is the first one in the wiki where the harness is the only variable.
It is also a partial reply to the same-day negative result. How Do Agents Fail on AutoResearch: End-to-End Diagnostic Evaluation on 100 Real-World Frontier Research Tasks (arXiv:2608.14905) finds failure patterns recurring across all 8 harness-model combinations and says outright that it does not test whether orchestration-level interventions can close the gap. ClawGym II is an orchestration-level intervention — but it closes the gap by changing the weights, which is the one move that leaves the model-level diagnosis intact rather than refuting it.
Practical note: the two named harnesses are shipped consumer products, not research scaffolds. Training against a vendor's agent harness makes the harness a dependency of the resulting weights, which nothing read discusses.
Open Questions
- Base rates. +9.98 and +14.81 points from an unstated starting point cannot be compared to any other agentic-RL result, and this wiki now has five mutually incomparable denominators in this cluster.
- Why is Claude Code worth 1.48× OpenClaw? The abstract states the difference and does not explain it. Whether it reflects harness quality, headroom at the base rate, or fit between harness and objective is unaddressed.
- What does mix-harness training buy? It is presented as a contribution and no result is given for it.
- Does a model trained through one harness transfer to another? The obvious question given the 1.48× spread, and not addressed in anything read. Contrast StateM, whose whole claim is a runbook transferring unchanged.
- Cost. 200–400 optimization steps of long-horizon agentic rollouts through sandboxed harnesses is the expensive path; no compute figure was read.
- Author list, affiliation, licence — unknown; the paper was not read.
Cite
ClawGym II: Exploring Black-Box RL on Agent Harness (2026). arXiv:2608.16798.
Related
- Eval Harness Configuration
- Agentic Reinforcement Learning
- Agents (LLM Agents)
- StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (arXiv:2608.15089)
- How Do Agents Fail on AutoResearch: End-to-End Diagnostic Evaluation on 100 Real-World Frontier Research Tasks (arXiv:2608.14905)