Meta’s Breakthrough: Training an 8B AI Model to Rival Claude Opus 4.5 at a Fraction of the Cost

Imagine an AI agent assigned the challenging task of transferring extensive customer data from an outdated CRM system to a cloud-based database. The agent’s internal context alone is not enough for this time-consuming operation, especially when it relies on a supportive runtime layer often referred to as the harness.

This harness is designed to offer real-time feedback, such as server logs, which help the agent effectively navigate changing API connections. It includes tools for tracking task completion and controlling the flow of processes, ensuring that the agent correctly handles all data batches without repeating or overlooking any. When unexpected issues arise, like a database rejecting a submission due to strict API limits, the harness assists in recovery with appropriate tools and guidance.

Typically, an agent is directed through a sequence of hardcoded rules written by human developers. For instance, a developer might specify that the agent should always check the company’s wiki prior to drafting an email. However, such rigid programming limits the agent’s ability to act independently, as it lacks the capacity to evaluate the consequences of its actions.

To address these limitations, researchers from Meta AI and the University of Illinois Urbana–Champaign have introduced EvoHarness-RL, a framework that enhances the agent’s harness by teaching it when and how to process information from its environment.

For AI agents engaged in long-term tasks, the capability to interpret and integrate environmental information is crucial for their success. They need to continuously update their understanding, track their progress on subgoals, recover from failures, and apply previously learned procedures. This is where the harness plays a vital role.

Previous frameworks, such as Harness-1, partially address this challenge by storing past experiences in a structured manner for use in future tasks. However, they often segregate long-term memory from real-time task management, leaving agents without sufficient guidance on how to handle immediate situational demands.

Xuying Ning, one of the co-authors of the EvoHarness-RL study, noted that the rigid structures for logic and memory are key factors that drain engineering resources.

“The optimal harness frequently varies with the model,” Ning said. “Various models may require distinct prompts, memory configurations, or permissions. Relying on manual coding means that every model update could necessitate an extensive tuning and debugging process.”

Moreover, existing memory systems that accumulate experiences can degrade the reasoning abilities of an agent. Ning explained, “An append-only memory assumes that having more context is always beneficial, which isn’t always true.” Over lengthy tasks, an agent’s memory may fill with outdated information and failed attempts, making it essential for long-horizon agents to have an adaptive memory that can refresh, condense, and replace outdated knowledge.

EvoHarness-RL: A Comprehensive System for Knowledge and Progress

To overcome the constraints imposed by static instructions, the researchers have introduced EvoHarness-RL, a training technique that enables the agent to make effective use of its harness. Rather than relying on fixed commands, the agent learns to create an organized workspace out of chaotic execution data and to decide when to utilize external information during intricate workflows.

To streamline the management of various harness components, EvoHarness-RL integrates the agent’s support mechanisms into a singular, cohesive interface known as Belief, Progress, and Experience (BPE), which breaks down into three key areas:

• Belief: Keep an accurate understanding of the current environment.

• Progress: Track completed objectives and outstanding tasks.

• Experience: Leverage past knowledge for new tasks.

Instead of engaging with complex, domain-specific APIs, the AI operates using a straightforward dashboard and four concise meta-actions: track, commit, recall, and note. These commands allow the agent to monitor real-time environments, commit to workflow updates, recall methods before acting, and take notes for future reference.

These components align directly with high-demand sectors. “In software engineering, Belief can represent the agent’s understanding of the repository,” noted Ning, explaining how the agent tracks interactions and changes in its workspace. Progress monitors what has been accomplished and what remains to be done, while Experience captures lessons learned from past errors to shape future actions.

The same framework can be applied in finance. During compliance checks, Belief might detail relevant regulations and evidence, Progress keeps tabs on completed checks and outstanding exceptions, and Experience assists the agent in recognizing frequent discrepancies and when escalation is warranted.

“These states work together to prevent the agent from losing its way or repeating previous mistakes,” added Ning.

To effectively teach agents how to manage their external workspace, the researchers devised a two-phase training regime. In the first phase, supervised harness fine-tuning, the fundamental model learns to extract useful insights from chaotic interaction logs into the BPE system.

However, accessing memory or updating trackers incurs time and resource costs, making it inefficient for the agent to continually check its tools. The second phase, termed “cost-aware” reinforcement learning, trains the agent to efficiently assess when it makes sense to access its external states. This two-phase approach allows for the evolution of tool usage from predefined prompts to a learned runtime behavior.

EvoHarness-RL in Practice

The efficacy of EvoHarness-RL was evaluated using the ALFWorld benchmark, a text-based setting designed to test tasks requiring sequential logic and state management.

The researchers employed the Qwen3-8B model for training and compared its performance against several leading models and static frameworks. The results indicated a remarkable improvement for smaller, budget-friendly models. The Qwen3-8B, powered by EvoHarness-RL, achieved an impressive 96.9% success rate, marking a notable 49.0 percentage point uplift from its baseline performance.

Additionally, the trained model outshone advanced training methods like SkillRL and SkillOS. Impressively, the model’s performance approached that of higher-end, resource-intensive models such as Claude Opus 4.5, which achieved a 96.4% success out of the box.

The experiments not only benefited smaller models but demonstrated the universal applicability of the BPE framework, enhancing the performance of all scales of models—even those without extensive reinforcement learning training. For instance, integrating the BPE prompt-time harness improved GPT-4.1 and GPT-5’s success rates significantly.

Further observations during the experiments revealed dynamic improvements in behavior as the LLMs navigated the EvoHarness-RL training. It was noted that the AI gradually reduced its dependence on querying its Experience and Progress trackers as it refined its understanding through knowledge accumulation, a process referred to as “harness annealing.”

Initially, the agent relied heavily on retrieving information for nearly every action but began to internalize successful strategies over time. In practical terms, this leads to lower operational latency and reduced computing costs, as the AI minimizes unnecessary database queries for routine tasks it has already mastered.

Simultaneously, the concept of “harness evolution” emerged, wherein the agent adapted its strategy in response to the complexity of the tasks at hand. While it tended to bypass its tools for straightforward tasks, it recognized the need to utilize its Belief and Experience modules when confronting novel challenges.

Integrating EvoHarness-RL into Current Systems

Implementing a new framework can often disrupt enterprise workflows. Yet, EvoHarness-RL includes an environment adapter that allows its integration while keeping existing tools and systems intact.

Ning expressed optimism, stating, “There’s considerable potential to incorporate BPE into current orchestration systems without replacing existing frameworks. BPE can function as an additional layer managing the agent’s beliefs, progress, and learned experiences.”

For enterprise developers concerned about costs associated with inference, the framework offers a solution to the hidden engineering expenses of consolidation. Given that consolidation requires robust reasoning capabilities, teams may consider employing a hybrid system to streamline expenses.

“A potential compromise is to leverage a high-end model for generating quality consolidation data, then fine-tune an open-weight model for routine state management,” explained Ning. Furthermore, “asynchronous consolidation can occur without delaying the main execution of the agent.”

Organizations must assess when a trainable BPE harness is essential as opposed to when simpler approaches might suffice.

“For quick and uncomplicated tasks, simpler frameworks like ReAct or standard RAG may already be adequate,” noted Ning. “BPE is particularly advantageous for agents engaged in lengthy, ongoing tasks.” In such intricate cases, agents require a condensing understanding of their decisions to prevent disorientation, leveraging Experience to improve over time based on previous outcomes and human input.

Overall, this evolution signals a significant development for AI orchestration engineers. “This isn’t about completely replacing workflow engineering,” Ning concluded, “but rather transitioning from directly coding agent behaviors to developing systems that enable better learning and adaptability.”