$ cat wiki/papers/2026/2608.17528-agent-lightning-v1.md
Agent Lightning v1.0: Towards Harnessed Agentic RL (arXiv:2608.17528)
TL;DR
Names and formalises harnessed agentic RL — the regime where the deploy-time harness owns the environment interaction loop and the trainer sees only sequences of LLM request/response pairs — and ships a ~3,500-line framework for it. Reported result: Qwen3.5-9B on SWE-bench Verified 41.8% → 56.4%, a 14.6-point absolute gain from 6K training examples and "modest compute" (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, 2026-08-20, 16 upvotes — that community's popularity signal and nothing more (source).
Method
The architecture is disaggregated: arbitrary agents connect to RL training through an LLM endpoint proxy, so the training system never has to understand the agent. The paper states this approach was introduced by the original Agent Lightning and later adopted by verl Uni-Agent, AReaL 2.0, slime and Polar — which, if accurate, makes the pattern the field's default rather than one option.
The paradigm's defining consequence, and the paper's own framing of what makes it hard:
| Traditional agentic RL | Harnessed agentic RL |
|---|---|
| The training engine owns the environment interaction loop | The harness owns it |
| The trainer observes the environment | The trainer observes only LLM request–response pairs |
| Stated open engineering problems that follow: retokenization, **sample | |
| merging**, advantage calculation, loss normalization, **backend | |
| scheduling** — each said to substantially affect training stability. |
Evaluated on instruction-following, search and coding agents, with a complete reproducible pipeline released for coding-agent RL.
Results
| Item | Figure |
|---|---|
| Model | Qwen3.5-9B |
| Benchmark | SWE-bench Verified |
| Before → after | 41.8% → 56.4% |
| Absolute gain | 14.6 points |
| Training examples | 6K |
| Framework size | ≈ 3,500 lines |
| What the abstract does not give: the harness used for the SWE-bench result, | |
| the compute behind "modest", any figure for the instruction-following or search | |
| agents, and any ablation separating the framework's engineering fixes from the RL | |
| itself. |
Significance
This is the harness cluster's training-side counterpart, and it inverts the cluster's usual variable. Every harness result this wiki holds treats the harness as a deployment-time choice measured against fixed weights — including StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (arXiv:2608.15089), whose entire claim is "weights untouched". Here the harness is present during post-training, so the weights are fitted to it.
That makes it the mechanism behind ClawGym II: Exploring Black-Box RL on Agent Harness (arXiv:2608.16798)'s most awkward finding, which this wiki recorded on 2026-08-19: after training through an opaque harness, the model is no longer harness-agnostic. ClawGym-II observed the effect (+9.98 via OpenClaw against +14.81 via Claude Code, same model and method — 1.48× apart); Agent Lightning describes the architecture that produces it and names the five places it goes wrong. Neither cites the other; the pairing is this wiki's.
The consequence for Eval Harness Configuration is concrete. That page's rule is that a benchmark number is a claim about a (model, harness) pair. If harnessed agentic RL is now the default training route — and the adoption list in the abstract argues it is — then the pair is not merely a reporting convention: it is baked into the weights, and a model's published score is partly a score for the harness it was trained against. A "harness-agnostic model" becomes a thing that has to be demonstrated rather than assumed.
Against that, the reported gain is the least interesting part. 41.8 → 56.4 on SWE-bench Verified from 6K examples is a large move for a 9B model, but the harness is unnamed, so by this page's own rule the figure is an incomplete claim.
Open Questions
- Which harness produced 56.4%? The paper's central thesis is that the harness participates in training; omitting its identity from the headline result is the exact gap Eval Harness Configuration exists to flag, in a paper arguing for the harness's importance.
- Does the gain transfer off that harness? ClawGym-II says it should not, fully. Nothing read tests it.
- "Modest compute" — unquantified, and the claim that ES-free RL is now cheap at 9B scale rests on it. Compare Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements (arXiv:2608.17310), which makes a competing efficiency claim by a different route.
- Is the adoption list independently verifiable? Four named frameworks are said to have adopted the paradigm; that is a strong priority claim made in an abstract.
- Author list, affiliation, licence — unknown; the paper was not read.
Cite
Agent Lightning v1.0: Towards Harnessed Agentic RL (2026). arXiv:2608.17528.
Related
- Agentic Reinforcement Learning
- Eval Harness Configuration
- ClawGym II: Exploring Black-Box RL on Agent Harness (arXiv:2608.16798)
- Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements (arXiv:2608.17310)
- StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (arXiv:2608.15089)