Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-04 · high
A separate harness engineer model rewrote a frozen 9B agent's runtime, raising task success from 44.3% to 53.6%.
Why now: The paper posted to arXiv in August 2026, with its 9.3-point gain tested so far only on WebShop, ALFWorld, and DBBench.
Harness-R1 lifts a frozen 9B Qwen agent's task success from 44.3% to 53.6% by training the runtime around it instead of the model, per a paper posted to arXiv.
That's a 9.3-point jump in task success without retraining the agent model at all. The gain shows up on WebShop, ALFWorld, and DBBench, benchmarks that test whether an agent actually finishes real tasks.
A separate harness engineer model watches the frozen target agent fail. It converts those failures into validated, executable patches to the harness. That's the code that builds context, mediates tool calls, validates actions, and recovers when execution breaks.
The harness engineer gets rewarded by rerunning the same batch of tasks against the frozen target agent after each patch lands. No new weights for the agent itself, just a better runtime wrapped around the same model.
Vanilla Qwen3.5-9B starts at 44.3% task success. Add the trained harness engineer and it climbs to 53.6%, a 9.3-point gain. Stack a second, target-specific harness engineer on top of the one already fine-tuned, and it adds another 5.0 points.
A one-time patch would be enough if the harness fix generalized cleanly. It doesn't. The extra 5 points from a target-specific engineer means the runtime around an agent needs the same per-deployment tuning prompts and fine-tunes already get. The paper doesn't say what a second training pass costs per deployed agent, or how the approach holds up past these three benchmarks.
Each link below shares sources, entities, or timing with this story.
SkillPyramid extends Voyager-style libraries with a hierarchical topology plus a self-evolution loop, raising average reward 38% and cutting execution steps 27.7% across ALFWorld, WebShop, and ScienceWorld. The takeaway: don't append every successful trajectory as a new flat s...
SkillForge (arXiv 2608.24747) notes that skill-extraction approaches like SkillRL never verify whether a stored skill still works against the current environment, so the bank grows monotonically while quality rots. It makes skill usage explicit during interaction so RL optimiz...
Policy-Aware Training Scaffolding converts rollout groups from the latest policy into evidence cards and adjusts agent context per task, removing guidance as the agent improves. The scaffold is discarded at deployment — the skills exist to shape training, not to ship. 18.6% ov...
The method turns tool use from a hardcoded prompt into a learned runtime behavior, then applies cost-aware RL teaching the agent when reading external state is worth the token budget. Qwen3-8B reaches a 96.9% average success rate against SkillOS at 80.2% and SkillRL at 89.9%,...
arXiv 2607.26598 targets the failure where an agent recovers from an error within an episode but hits the identical failure in later tasks, because post-episode feedback never revises the persistent harness. Guided by a domain-level Evolution-SOP, it writes episodic memory rec...
Repo-To-Skill (arXiv 2609.02749, submitted September 2) makes the argument I've been circling for months and then supplies the controlled experiment. The claim is that what research agents lack is operational know-how, not reasoning capability, and that this know-how already e...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.