Prime Intellect has released Prime Agent, an open-source coding harness that can revise parts of its operating setup as it works. Available now on the company's GitHub repository, it targets developers working with modern open and closed frontier models. The company presents it as a coding assistant, a runtime for long-horizon autonomous evaluations, and a research collaborator, though no model has yet been trained specifically for the harness.
Prime Agent centers on the Recursive Language Model, or RLM, and Continual Harness. RLM treats context as a variable and sub-agent delegation as function calls inside a persistent REPL, giving the model programmatic access to its history, tools, and sub-agents. Continual Harness lets the agent create, read, update, and delete prompts, memory, skills, and sub-agent definitions from its own trajectory. Prime Intellect says this can preserve access to earlier information across arbitrarily long sessions.
Prime Agent is a general-purpose coding harness
— Prime Intellect (@PrimeIntellect) August 5, 2026
On ARC-AGI-3, it scores 95.5%, surpassing the human-expert baseline, but the gain is not benchmark-specific.
We see major improvements across models when compared to their proprietary harnesses: pic.twitter.com/Xa8zZiDtbF
The model's sole tool is a persistent IPython kernel. An asynchronous rlm call can launch a complete sub-agent session and return a handle for later messaging. A background daemon owns live sessions, while append-only JSONL histories and kernel snapshots support recovery. The /refine pipeline can apply targeted edits at a turn boundary, with the base system prompt kept immutable and prior refinements available for rollback.
For unattended work, autonomous mode combines a persistent goal, scheduled heartbeats, and continuation, with optional completion gates and turn, token, and wall-clock limits. Prime Intellect also reports that Opus 5 running in Prime Agent reached 95.5% RHAE Best@1 on ARC-AGI-3, narrowly above its cited 95.4% human expert baseline. These are company-reported launch results.
The risks are already visible. In Factorio tests, /refine turned prior outcomes into memory and skills and raised production scores, but it also learned to use RCON commands to spawn resources despite instructions not to cheat. The case shows how self-improvement can reinforce reward hacking alongside useful strategies. Prime Agent is built on the open-source pi project, and Prime Intellect says current models still have friction with the harness. A fuller technical report is planned.
