49 minutes ago
17
Researchers from Meta AI and the University of Illinois Urbana-Champaign have published a new paper introducing EvoHarness-RL, a trainable coordination layer that lets large language model agents create, access, and manage their own external state. The result: a 96.9% success rate on a standard benchmark, up from 47.9% for the baseline model running without it.
What EvoHarness-RL actually does
LLM agents have limited context windows. When a task stretches across many steps, involving dynamic API connections, server logs, pending subgoals, and error recovery, the model’s internal memory isn’t enough. It needs an external scaffold to track what it believes about the world, what progress it’s made, and what it’s learned from past mistakes.
The framework introduces a structured external state built around three components: Belief, Progress, and Experience, collectively called BPE. Belief captures the agent’s current understanding of its environment. Progress tracks completed and pending subgoals so the agent doesn’t skip steps or duplicate work. Experience stores lessons from errors, like a database rejecting a batch due to API rate limits, so the agent can adapt its approach.
Training happens in two stages. First, supervised fine-tuning using expert demonstrations teaches the model how to interact with the harness. Then a cost-aware optimization technique called Group Relative Policy Optimization, or GRPO, refines the agent’s behavior to balance task performance against the computational cost of harness calls.
The researchers tested their approach using the Qwen3-8B model on ALFWorld, a benchmark for embodied AI that requires agents to complete household tasks across multiple steps. The 96.9% success rate in familiar environments is impressive on its own, but perhaps more telling is the 86.6% success rate in unseen environments.
Two behaviors the researchers didn’t explicitly program
The paper highlights two emergent phenomena that surfaced during training, neither of which was directly engineered.
The first is “harness annealing.” Over time, the agent begins to internalize routine harness operations, reducing how frequently it needs to consult the external state. The harness calls decrease not because the system is degrading, but because the model has absorbed the patterns.
The second is “harness evolution.” As the agent trains, it learns to compress and restructure its external state into a more compact format suited for specific task types. The BPE representation becomes leaner and more specialized without any human intervention in its design.
Building on earlier work
EvoHarness-RL builds on Meta-Harness, an earlier project from March 2026 that focused on optimizing harness code through agentic search techniques. Where Meta-Harness treated the harness as code to be improved through search, EvoHarness-RL treats the entire coordination layer as something the model itself can learn to construct and refine.
The paper also reports improvements over existing agent frameworks like SkillOS and SkillRL, particularly on unseen tasks. The gap between performance on familiar versus novel environments is roughly 10 percentage points for EvoHarness-RL, compared to much larger drops for competing approaches.
What this means for enterprise AI deployment
A success rate jumping from roughly 48% to 97% in controlled environments isn’t a marginal improvement. The 86.6% generalization rate matters too, because real-world enterprise tasks rarely look exactly like training data.
The cost-aware optimization built into the GRPO training stage also addresses a practical concern. External harness calls aren’t free. They consume compute and add latency. By training the agent to minimize unnecessary calls through harness annealing, the framework becomes more efficient over time without sacrificing accuracy.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·