← Back to feed

A paper published this week on arXiv — Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses (arXiv:2606.02373) — describes a 20B model that outperforms the next-strongest open search agent by 11.4 points on a curated recall benchmark, and remains competitive with much larger frontier-model searchers.

The mechanism is architectural rather than computational: the paper separates what the model does from what the environment tracks, arguing that most production search agents have been training models to solve two fundamentally different problems at once without naming either of them.

The Problem

Standard search agents maintain everything in a growing context window. As a session extends, the model must simultaneously make semantic search decisions — what to query next, what evidence to trust, when to stop — and track administrative state: the candidate pool, prior search paths, verification records, curated documents, and evidence already seen.

The Harness-1 authors call this formulation “putting too much routine state management inside the policy.” The consequences are measurable. RL training is particularly sensitive to the problem: when an agent fails to find the right answer, the outcome looks the same whether the agent searched badly, forgot earlier evidence, skipped verification, or curated carelessly. Bad reward signals degrade learning efficiency across the board.

The Split

Harness-1 separates the responsibilities cleanly:

The policy model handles:

• What to search
• What to inspect
• What to curate and verify
• When the evidence is sufficient to stop

The harness handles:

• Candidate pools
• Curated document lists
• Importance tags
• Evidence links
• Verification records
• Deduplicated observations
• Budget-aware memory rendering (deciding what to surface in context, and when)

The model never has to remember whether it already tried a specific search path. The harness keeps that. The model’s context stays focused on decisions that require semantic judgment.

Results Across 8 Benchmarks

Harness-1 reaches 0.730 average curated recall across 8 difficult retrieval benchmarks. The next-strongest open search subagent trails by 11.4 points. The system remains competitive with frontier-model searchers built on models that are 5-10x larger.

The most significant finding is the transfer result: gains on held-out benchmarks are substantially larger than gains on source-family benchmarks. This is the signal that the model learned generalizable search behaviors rather than memorizing domain-specific patterns. Ablations confirm it — removing the harness mechanisms changes agent behavior and degrades recall, ruling out the possibility that the policy was doing the same work internally.

Why This Matters for Production Agents

The paper’s framing has direct implications for how agent infrastructure is designed. The standard approach treats the LLM as the complete learning system and adds scaffolding to handle failures. Harness-1 argues the harness is a first-class component — “part of what RL learns to use” — and should be designed explicitly rather than bolted on.

For retrieval-augmented generation pipelines, the implication is that improving search quality may not require a larger base model. It may require a better-designed harness that separates recoverable state from genuine reasoning so the model learns the right thing.

The architectural pattern also connects to a broader theme in recent agent research: two papers published in May — covered previously — found that a single strong model outperforms multi-agent pipelines on many tasks, and that grep-based retrieval beats embedding search on structured codebases. Harness-1 sits in the same space: the gains come from removing overhead that was silently degrading performance, not from scaling the core model.

The paper is open access on arXiv at 2606.02373.