Prime Agent Rewrites Its Own Harness While Running
Prime Intellect released Prime Agent on August 5 (primeintellect.ai/blog/prime-agent): an open-source coding and research agent built around one idea — the agent should be allowed to edit itself. Not its weights. Its harness.
Two abstractions do the work. First, the Recursive Language Model: instead of one context window that fills up and dies, the agent treats context as dynamic and delegates subtasks as function calls inside a persistent IPython REPL, writing what the team calls "language model programs as actions over its own context." Second, the continual harness: the agent can create, read, update, and delete its own prompts, skills, memory, and sub-agents while it runs, with a /refine pipeline that turns observed failures into targeted harness improvements. Agents run as a family tree — parent, sibling, child — coordinated by a background daemon.
The headline number: 95.5% on ARC-AGI 3, a hair above the human expert baseline of 95.4%. Plus competitive long-context results against closed-model harnesses, and demos ranging from building emulators to writing GPU kernels. Take single-benchmark claims with the usual salt, but crossing a human-expert line on ARC with an open-source harness is not nothing.
This is the harness-beats-weights thread reaching its logical endpoint. LongHorizon-Harness showed this week that external task-state can add 30 points to a fixed model. SkillRise argued skills should accumulate across tasks. Prime Agent closes the loop: if the harness is where the gains are, let the agent optimize the harness itself. Whether /refine produces real compounding improvement or just churn is the question that matters — and because it is fully open source, anyone can go check.
← Back to all articles
Two abstractions do the work. First, the Recursive Language Model: instead of one context window that fills up and dies, the agent treats context as dynamic and delegates subtasks as function calls inside a persistent IPython REPL, writing what the team calls "language model programs as actions over its own context." Second, the continual harness: the agent can create, read, update, and delete its own prompts, skills, memory, and sub-agents while it runs, with a /refine pipeline that turns observed failures into targeted harness improvements. Agents run as a family tree — parent, sibling, child — coordinated by a background daemon.
The headline number: 95.5% on ARC-AGI 3, a hair above the human expert baseline of 95.4%. Plus competitive long-context results against closed-model harnesses, and demos ranging from building emulators to writing GPU kernels. Take single-benchmark claims with the usual salt, but crossing a human-expert line on ARC with an open-source harness is not nothing.
This is the harness-beats-weights thread reaching its logical endpoint. LongHorizon-Harness showed this week that external task-state can add 30 points to a fixed model. SkillRise argued skills should accumulate across tasks. Prime Agent closes the loop: if the harness is where the gains are, let the agent optimize the harness itself. Whether /refine produces real compounding improvement or just churn is the question that matters — and because it is fully open source, anyone can go check.
Comments