Prime Agent: A self-improving RLM agent
Summary
Prime Agent uses Recursive Language Model (RLM) and Continual Harness abstractions to achieve 95.5% on ARC-AGI-3, surpassing the human expert baseline.
RLM treats context as a variable.
The Recursive Language Model uses a persistent REPL to allow programmatic access to history and tools, enabling agents to write programs over their own context.
treats context as a variable and subagent delegation as function calls
Continual Harness enables agent self-modification.
The harness allows agents to create, read, update, and delete their own prompts, skills, memory, and sub-agents during a trajectory.
Opus 5 beats human experts on ARC-AGI-3.
Using Opus 5, Prime Agent reached 95.5% RHAE Best@1, exceeding the 95.4% human expert baseline.
surpasses the ARC reported human expert baseline of 95.4%
Refinement loops improve game production scores.
In Factorio, Prime Agent used the `/refine` command to convert experiences into memories and skills, reaching production scores over 100K within hours.
leveraged /refine to turn failures and successes into memories and skills
Direct training could unlock more performance.
The authors suggest that current friction between models and the harness implies significant gains remain if models are trained directly with the RLM and Continual Harness paradigms.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.