Prime Agent Builds a Persistent Runtime for Long-Horizon Agents

Introduction

Long-horizon agents often fail for reasons that are difficult to attribute to the underlying model. Context may grow beyond a useful working window, intermediate computation may be lost, a tool call may leave the environment in an unclear state, or a recovery script may simply break. In such cases, the benchmark score reflects the fragility of the surrounding harness as much as the model’s reasoning ability. Prime Agent is an open-source attempt to make that surrounding layer persistent, inspectable, and easier to improve.

What the harness provides

• A persistent IPython REPL: Prime Agent follows the Recursive Language Model, or RLM, abstraction. The model can process context programmatically and use additional test-time computation. Because the REPL persists, useful variables, outputs, and intermediate work do not have to be reconstructed after every interaction.
• Continual state across trajectories: Its Continual Harness stores histories, memories, skills, prompts, and specifications for subagents. This gives an agent a way to carry reusable experience into later runs instead of treating every trajectory as a clean slate.
• Recursive agent-to-agent communication: Subagents can communicate directly and divide a task among specialized workers. The design is suited to workflows in which some agents explore, others implement, and additional agents verify or refine the result.
• Execution and recovery as standard services: Prime Agent standardizes execution, failure recovery, verification, and resource accounting. The Agents View gives humans a way to inspect and manage daemon-backed sessions, making long-running processes less opaque.

Reported evaluation

According to the paper, Prime Agent raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5%. It also matches or exceeds native and popular harnesses on long-context coding, GPU-kernel generation, emulator construction, and autonomous nanoGPT speedruns. In Factorio, the authors report that iterative refinement supports continued technological progress, while dedicated subagents make parallel work possible.

The broader lesson is that an agent benchmark is partly a runtime benchmark. A system that can retain state, call computation, recover from errors, and verify its own work may expose capabilities that remain hidden behind a less capable harness. At the same time, the abstract does not include the full experimental configurations, cost breakdowns, or ablation results. The reported gains should therefore be read as results for the complete agent-and-harness system, not automatically as equivalent gains in the base model alone.

Why it matters

Prime Agent deliberately leaves strategy construction to the model. Its role is to provide a low-friction membrane through which the model can create and refine its own workflow without being derailed by routine infrastructure problems. For researchers, this can move evaluation closer to a model’s practical maximum under a robust runtime. For developers, persistent sessions, memory, verification, and accounting are reusable building blocks for coding agents that operate over extended periods.

This points to a broader shift in agent engineering. Future progress may depend not only on larger models, but also on better runtimes, memory policies, coordination protocols, and recovery mechanisms. Prime Agent offers an open implementation for exploring that direction. Its generality, cost profile, and benefits across different models still require independent reproduction.

Source: Hugging Face Daily Papers