saved
Prime Agent: A Self-Improving RLM Harness
gist
Prime Agent is an open-source harness for long-horizon evaluation and coding-agent work. A persistent IPython REPL uses the Recursive Language Model abstraction for programmatic context and test-time compute. Continual Harness keeps histories, memories, skills, prompts, and subagent specifications across trajectories. Recursive subagents coordinate directly, and an Agents View lets humans inspect daemon-backed sessions. The harness standardizes execution, recovery, verification, and accounting so failures measure the model, not dropped state. It reports ARC-AGI-3 RHAE Best@1 from 30% to 95.5%.
ideas
- Treat the harness as a membrane. Standardize execution, recovery, verification, and resource accounting so a model fails because the task exceeds it, not because the harness dropped state.
- Make context programmable. A persistent IPython REPL follows the Recursive Language Model abstraction for programmatic context processing and test-time compute.
- Keep harness state across trajectories. Continual Harness preserves histories, memories, skills, prompts, and subagent specifications.
- Coordinate recursive subagents. Direct agent-to-agent communication plus an Agents View for inspecting daemon-backed sessions.
- Reported lift is largest on interactive reasoning. ARC-AGI-3 RHAE Best@1 rises from 30% to 95.5%, and the abstract says the harness matches or exceeds native ones on long-context coding, kernels, emulators, and nanoGPT speedruns.
quotes
“Language models are sequential processors, but long-horizon agency requires external information and computation beyond model weights and active context.”
“Prime Agent is an open-source harness for long-horizon evaluation and coding-agent workflows.”
“This low-friction, expressive membrane prevents harness failures from becoming model failures”
“Prime Agent raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5%”