Meta AI researchers, working with the University of Illinois Urbana - Champaign, have developed a training framework called EvoHarness-RL that enables a compact 8-billion-parameter AI model to match the performance of much larger frontier systems on complex, long-horizon tasks. The work targets a core bottleneck in enterprise AI: teaching agents to manage their external tools and memory efficiently rather than relying on rigid, human-coded instructions. Using the ALFWorld benchmark, the trained model reached a 96.9% average success rate, edging out Claude Opus 4.5, which scored 96.4% out of the box.
Long-running workflows, such as migrating customer records between systems, require an agent to track progress, recover from errors, and update its understanding of a changing environment. The harness, or runtime layer, provides these capabilities, but its logic is usually hardcoded by developers, limiting autonomy. Co-author Xuying Ning said that manual logic drains engineering resources because the optimal harness often changes with each model, leading to lengthy tuning cycles when upgrades occur.
EvoHarness-RL addresses this by giving the agent a unified interface called Belief, Progress, and Experience, or BPE, which consolidates its external support systems. The model interacts with this dashboard using four meta-actions: track, commit, recall, and note. In software engineering, Belief represents the agent鈥檚 understanding of a repository, Progress tracks completed and pending steps, and Experience stores lessons such as user feedback. The same structure applies to finance, where Belief covers rules and evidence during a compliance audit, Progress lists checks to finish, and Experience flags recurring discrepancies.
The training process uses two stages. First, supervised fine-tuning teaches the base model to extract useful facts from messy interaction logs into the BPE framework. Second, cost-aware reinforcement learning trains the agent to decide when accessing its external state is worth the compute budget, avoiding unnecessary queries. This transforms tool use from a hardcoded prompt into a learned behavior.
In tests, the 8-billion-parameter Qwen3-8B model, trained with EvoHarness-RL, outperformed other trainable frameworks such as SkillRL, which scored 89.9%, and SkillOS, which scored 80.2%. The team also found that the BPE framework improves frozen frontier models without the Reinforcement Learning phase. GPT-4.1鈥檚 success rate rose by 22.1 points, and GPT-5鈥檚 by 25.7 points, when equipped with the prompt-time harness.
Researchers observed a behavioral shift during training they call "harness annealing." Early on, the agent queried its Experience and Progress trackers at nearly every step, but over time it internalized routine actions and relied less on external checks. For US enterprise developers seeking to cut compute costs, the results suggest smaller models can match expensive closed systems on demanding tasks without sacrificing accuracy.
More AI news from TechManNews.
