Prime Agent is an open-source harness for long-horizon evaluation and coding-agent workflows. It uses a persistent IPython REPL following the Recursive Language Model abstraction for programmatic context processing and test-time compute. Continual Harness preserves histories, memories, skills, prompts, and subagent specifications across trajectories. Recursive subagents coordinate through direct agent-to-agent communication. Agents View lets humans inspect and manage daemon-backed sessions. Prime Agent standardizes execution, recovery, verification, and resource accounting. It raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5% and matches or exceeds native and popular harnesses across long-context coding, GPU-kernel generation, emulator construction, and autonomous nanoGPT speedruns.
Prime Agent introduces a persistent IPython REPL as a Recursive Language Model abstraction, enabling programmatic context processing and test-time compute. The Continual Harness persists state across trajectories, allowing models to retain memories, skills, and subagent specifications. Recursive subagents communicate directly, enabling hierarchical coordination. The harness standardizes execution, recovery, verification, and resource accounting, reducing infrastructure-induced failures. The reported improvement on ARC-AGI-3 RHAE Best@1 from 30% to 95.5% suggests significant gains in long-horizon reasoning and agentic task performance.
Prime Agent addresses a key bottleneck in AI agent development: the harness itself can limit model performance. By providing a low-friction, expressive membrane, it allows models to demonstrate their true capabilities. This could accelerate adoption of agentic workflows in software development and other long-horizon tasks. The open-source nature may foster a community of developers building on the harness, potentially leading to standardized evaluation and deployment practices.
Prime Agent reduces the engineering overhead for building and evaluating long-horizon AI agents. By standardizing execution and recovery, it can lower development costs and time-to-market for agentic products. The performance gains on benchmarks may translate to more capable coding assistants and autonomous systems, creating value for enterprises seeking to automate complex workflows.
Next signals to watch include adoption of Prime Agent in benchmark leaderboards, integration with popular coding agents, and community contributions extending the harness. If the reported ARC-AGI-3 improvements are replicated, it may influence evaluation standards for long-horizon agency. Potential commercial applications include autonomous software engineering and complex task automation.