Prime Agent: A Self-Improving RLM Harness

TL;DR

Prime Agent is an open-source framework that enables long-horizon AI agent tasks by providing persistent execution environments, memory retention, and coordination between subagents. It improves performance on complex reasoning and coding benchmarks by decoupling execution infrastructure from model strategy, allowing models to focus on planning while the harness handles recovery, verification, and resource management.

Why It Matters

The framework reduces the risk of system failures derailing agent performance and enables more reliable evaluation of AI capabilities in real-world scenarios. By supporting persistent sessions and parallelized workflows, it lowers the barrier to building and testing autonomous agents for complex, long-term tasks.

Key Findings

• Prime Agent increases ARC-AGI-3 RHAE Best@1 score from 30% to 95.5% through improved execution resilience and memory continuity.
• It enables autonomous nanoGPT speedruns, GPU kernel generation, and emulator construction by maintaining persistent programmatic context and test-time compute.
• Subagent coordination via direct communication and dedicated roles allows parallelized work and continuous technology progression in environments like Factorio.