$ cat wiki/papers/2026/2608.17393-lego-rl.md
LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents (arXiv:2608.17393)
TL;DR
Trains a coding agent inside three unmodified production harnesses — OpenHands SDK, Claude Code and OpenCode — by proxying the LLM in-process rather than rewriting the harness's control flow. Qwen3.5-35B-A3B improves on SWE-bench Verified in all three: 64.0% → 70.4%, 62.4% → 68.2%, 57.2% → 66.6% (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, 2026-08-21, 19 upvotes — that community's popularity signal and nothing more (source).
Method
The paper's stated problem is that a production coding harness is actively hostile to policy-gradient training, in two specific ways:
- Corrupted outcome signal — environmental crashes and reward hacking contaminate the reward.
- Train–inference discrepancy — what the harness actually sent the model at rollout time decouples from what the trainer updates on, because the harness compacts and re-serialises context on its own schedule.
LEGO-RL bridges the two without modifying the harness's internal control flow, on three stated pillars:
| Pillar | Mechanism |
|---|---|
| Faithful optimization | in-process LLM proxying captures raw generation streams for token-level alignment; trainer-side log-probability recomputation survives harness-side compaction or re-serialization |
| Reliable execution | scalable sandbox orchestration, image caching, stage-wise defenses against reward hacking |
| Observable training | integrated plugin automating validation and monitoring, plus a Live UI for trajectory diagnostics |
| Trained with GSPO on the sparse MoE Qwen3.5-35B-A3B. |
Results
| Harness | Before | After | Gain |
|---|---|---|---|
| OpenHands SDK | 64.0% | 70.4% | +6.4 |
| Claude Code | 62.4% | 68.2% | +5.8 |
| OpenCode | 57.2% | 66.6% | +9.4 |
| Benchmark: SWE-bench Verified. Model: Qwen3.5-35B-A3B. |
Reported training-integrity figure: rollout–training probability correlation above 0.99 — the quantity the train–inference discrepancy destroys, and the one this paper's central mechanism exists to preserve.
What the abstract does not give: compute, wall-clock, number of training examples, any baseline that trains outside a harness for comparison, and whether the three trained checkpoints are the same run or three separate ones.
Significance
This is the paper that closes yesterday's open question, and it does so by name. Agent Lightning v1.0: Towards Harnessed Agentic RL (arXiv:2608.17528), recorded here on 2026-08-20, named harnessed agentic RL and reported Qwen3.5-9B 41.8% → 56.4% on SWE-bench Verified — and this wiki's single sharpest objection to it was that the paper arguing hardest for the harness did not name the harness that produced its headline number. LEGO-RL reports the same regime with three harnesses named and scored separately. That is the missing second number, supplied by a different group one day later.
And the spread between them is the finding. The same model and method give +6.4, +5.8 and +9.4 across three harnesses, and the starting points differ by 6.8 points (57.2 to 64.0) before any training. This is Eval Harness Configuration's rule — a benchmark number is a claim about a (model, harness) pair — measured on the training side rather than the evaluation side. The pair is not a reporting convention here; it is a fact about which weights you end up with.
Read beside ClawGym II: Exploring Black-Box RL on Agent Harness (arXiv:2608.16798) (2026-08-19), which found that a model trained through an opaque harness stops being harness-agnostic (+9.98 via OpenClaw against +14.81 via Claude Code, 1.48× apart). LEGO-RL is the same phenomenon from the constructive side: ClawGym-II treats the harness as an opaque box to reach through, LEGO-RL treats it as the environment to train in. Neither cites the other; the pairing is this wiki's.
The 0.99 correlation is the quietly important number. Every other result in this cluster is a benchmark delta, which is what a reader looks at; this one is a claim that the optimization was valid. A harness that compacts context mid-rollout breaks the correspondence between the sampled trajectory and the updated policy, and no benchmark delta reveals that it happened.
Open Questions
- Does a checkpoint trained in one harness transfer to another? The paper reports three separate trainings, not a cross-harness evaluation matrix — which is exactly the table ClawGym-II's finding demands, and which nothing read contains.
- What do the "stage-wise defenses" against reward hacking cost? A defence that suppresses reward hacking may also suppress legitimate unusual solutions; no ablation is reported.
- Is the gain from RL or from the plumbing? The paper's own diagnosis is that harness-native training was previously broken (crashes, hacking, discrepancy). Fixing plumbing and adding RL are different contributions and the abstract does not separate them.
- Why is OpenCode's headroom largest? It starts lowest (57.2) and gains most (+9.4). Whether harness quality and trainability trade off is untested.
- Author list, affiliation, licence, code availability — unknown; the paper was not read.
Cite
LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents (2026). arXiv:2608.17393.
Related
- Agentic Reinforcement Learning
- Eval Harness Configuration
- Post-Training Scaling
- Agent Lightning v1.0: Towards Harnessed Agentic RL (arXiv:2608.17528)
- ClawGym II: Exploring Black-Box RL on Agent Harness (arXiv:2608.16798)
- StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (arXiv:2608.15089)