# Harness Evolution

RSI-Env places model training and harness modification in one persistent sandbox, with an independent Host Controller handling services, validity checks, versioning and capability evaluation, and in Self mode each accepted model–harness version takes over the next improvement round. Starting from Qwen3.8-27B with a minimal harness, Open-General Self, which receives no benchmark feedback, raises GPQA Diamond accuracy from 65.2% to 73.2% and AIME 2026 from 53.3% to 56.7%; Knowledge Self lifts SimpleQA from 82.0% to 85.0% and then 88.0%, while the same runs fall back on other benchmarks.

RSI-Env places model training and harness modification in one persistent sandbox, with an independent Host Controller handling services, validity checks, versioning and capability evaluation, and in Self mode each accepted model–harness version takes over the next improvement round. Starting from Qwen3.8-27B with a minimal harness, Open-General Self, which receives no benchmark feedback, raises GPQA Diamond accuracy from 65.2% to 73.2% and AIME 2026 from 53.3% to 56.7%; Knowledge Self lifts SimpleQA from 82.0% to 85.0% and then 88.0%, while the same runs fall back on other benchmarks.

An agent's performance depends not only on its underlying model but on its runtime harness: prompts, memory, skills, tools and orchestration. Harness2 makes harness improvement a two-level recursion. At the task level, a contrastive propose-probe-compose step compares execution behavior before and after edits without ground-truth answers; at the harnessing level, that process is distilled into an experience library that guides later tasks. On WorkBuddy-Bench, JobBench and Harvey's Legal Agent Benchmark, a single step improved both GLM-5 and Gemini 3.5 Flash across four model-harness configurations, by up to 11.5 points, 4.5 points and 11.8 percentage points.

An agent's performance depends not only on its underlying model but on its runtime harness: prompts, memory, skills, tools and orchestration. Harness2 makes harness improvement a two-level recursion. At the task level, a contrastive propose-probe-compose step compares execution behavior before and after edits without ground-truth answers; at the harnessing level, that process is distilled into an experience library that guides later tasks. On WorkBuddy-Bench, JobBench and Harvey's Legal Agent Benchmark, a single step improved both GLM-5 and Gemini 3.5 Flash across four model-harness configurations, by up to 11.5 points, 4.5 points and 11.8 percentage points.

Cortex turns each developer correction into a replayable task and keeps a skill or rule only when repeated paired comparisons show a net gain and no task collapses; across four independent runs it raised pass rates on four project conventions from 21.5% to 76.5% on scenarios no session had played.

Cortex turns each developer correction into a replayable task and keeps a skill or rule only when repeated paired comparisons show a net gain and no task collapses; across four independent runs it raised pass rates on four project conventions from 21.5% to 76.5% on scenarios no session had played.

The work introduces CoTrace, a harness-aware data recipe for alternating model–harness co-evolution that explicitly governs trajectory routing, provenance matching, and curriculum refresh: execution failures are routed to harness synthesis while policy supervision uses only verified successes matched to the adopted runtime. On the 102-task Tmax promotion split, Qwen3.5-9B advances from 78 to 88 solved tasks under supervised fine-tuning and reaches 90 with an online reinforcement variant, whereas much larger corpora pooled across sibling harnesses yield no accepted model update.

The work introduces CoTrace, a harness-aware data recipe for alternating model–harness co-evolution that explicitly governs trajectory routing, provenance matching, and curriculum refresh: execution failures are routed to harness synthesis while policy supervision uses only verified successes matched to the adopted runtime. On the 102-task Tmax promotion split, Qwen3.5-9B advances from 78 to 88 solved tasks under supervised fine-tuning and reaches 90 with an online reinforcement variant, whereas much larger corpora pooled across sibling harnesses yield no accepted model update.

EmbodiedRSI is a self-evolving agentic harness that, around a frozen robot foundation model, maintains code and skill hypotheses in a Hypothesis Graph, uses Value-of-Information Experiment Selection to pick the physical experiments that best distinguish them, and turns each interaction into reusable improvement through Code-Skill Co-Evolution and Reward-Grounded Memory Learning; it reaches 77.0% overall success on RoboCasa365 and 71.3% on Composite-Unseen (best baseline 40.1%), 86.8% overall on LIBERO-Pro, and transfers the simulation-evolved harness zero-shot to a real robot at 71.3% overall success.

EmbodiedRSI is a self-evolving agentic harness that, around a frozen robot foundation model, maintains code and skill hypotheses in a Hypothesis Graph, uses Value-of-Information Experiment Selection to pick the physical experiments that best distinguish them, and turns each interaction into reusable improvement through Code-Skill Co-Evolution and Reward-Grounded Memory Learning; it reaches 77.0% overall success on RoboCasa365 and 71.3% on Composite-Unseen (best baseline 40.1%), 86.8% overall on LIBERO-Pro, and transfers the simulation-evolved harness zero-shot to a real robot at 71.3% overall success.

HASTE is a multi-agent framework that evolves agent harnesses from sparse threat evidence, such as brief descriptions or a few attack examples in threat reports and preprints, through an adversarial interplay between safety-specification generation and attack-case generation, feeding evaluation outcomes back into both processes so harnesses evolve against emerging attacks beyond the initially observed evidence; experiments across multiple backbone models, attack types, and evidence forms show it consistently reduces attack success rates while preserving benign-task utility.

HASTE is a multi-agent framework that evolves agent harnesses from sparse threat evidence, such as brief descriptions or a few attack examples in threat reports and preprints, through an adversarial interplay between safety-specification generation and attack-case generation, feeding evaluation outcomes back into both processes so harnesses evolve against emerging attacks beyond the initially observed evidence; experiments across multiple backbone models, attack types, and evidence forms show it consistently reduces attack success rates while preserving benign-task utility.