Meta AI researchers have introduced a new method to help artificial intelligence agents handle complex, long-running workflows more efficiently. The system, called EvoHarness‑RL, is designed to improve how AI models manage their tools and memory during tasks that span hours or days, rather than just minutes.
Why agents struggle with long tasks
When an AI agent tries to move massive batches of customer records from a legacy system to a cloud database, it needs more than just internal instructions. The runtime layer, often called a harness, provides feedback and tracks progress.
Related: Cohere improves document parsing efficiency
If it cannot see the server logs or manage pending subgoals, data batches may be skipped or actions fail when rate limits block them. Standard setups usually force a human developer to write rigid rules, limiting the ability to learn from mistakes or adjust to changing conditions.
Meta AI and the University of Illinois Urbana–Champaign wanted a way to let agents update, compress, and replace information dynamically so they do not repeat the same errors. They built on earlier work such as Harness‑1 and created a unified workspace that consolidates support systems.
This interface, known as the Belief, Progress, and Experience (BPE) framework, divides external needs into three clear areas. Belief tracks current understanding of the environment, Progress manages completed and pending subgoals, and Experience stores historical knowledge for reuse.
Related: Visa deploys AI to fix code flaws
Training the agent to save money
The researchers introduce EvoHarness‑RL, a training technique that teaches the agent to make optimal use of its harness. Instead of blindly following hard‑coded instructions, the agent learns how to construct a structured workspace from messy execution data and decide when and how to consult that external state during complex workflows.
Practical advice for enterprise builders
For teams concerned about inference costs, a hybrid architecture might work well. A powerful frontier model can generate consolidation data, which is then used to fine‑tune a smaller, open‑weight model for routine state management. This allows the consolidation process to happen asynchronously, so it does not slow down the main execution loop.
