Prime Intellect releases Prime Agent, an open source harness for autonomous agents
On 5 August Prime Intellect published Prime Agent, an open source harness designed to run coding agents and long-horizon autonomous tasks. The code sits in the PrimeIntellect-ai/prime-agent repository under an MIT licence, and the announcement promises installation with a single command. It introduces an architecture built on the Recursive Language Model (RLM) and on a formalised "Continual Harness". Unlike static setups where "sub-agents, prompts, skills and memory... do not adapt to what the agent learns while it is running", this system uses a persistent IPython kernel in which sub-agents behave like pre-imported modules, handling state through CRUD operations and a background daemon.
According to the official announcement, Prime Agent reached 95.5% on the RHAE Best@1 metric of ARC-AGI-3 using the Opus 5 model, one tenth of a point above the human baseline of 95.4%. The three runs reported come in at 95.0%, 95.2% and 95.5%: the figure quoted is therefore the best of the three. It should be said plainly that these are vendor-supplied numbers, with no independent validation from the ARC Prize Foundation. Prime Intellect reports its own harness ahead of the proprietary tools bundled with each model in 8 evaluations out of 9 with GLM-5.2, and in 6 out of 9 with both Opus 5 and GPT-5.6 Sol, across a suite of nine long-context tests, and says it obtains those results while consuming fewer tokens overall than each model's native harness.
The release comes with case studies ranging from writing Rust emulators for retro consoles to running automated games of Factorio. The documentation, however, carries an explicit safety warning: the environment is not a "security sandbox", and its use is recommended on throwaway clones or restricted environments. The release puts the spotlight on something often overlooked: how well agents perform depends more and more on the software engineering wrapped around them, rather than on the raw power of the underlying model alone.
— Olya In an era obsessed with model parameters, it is reassuring to see someone focus on execution architecture; a shame that checking whether the harness really keeps its promises means trusting the estimates of the people selling it.
Come Olya ha verificato questa notizia
- Verificato
- Read the official announcement on the Prime Intellect blog (primary source) for the architecture, the ARC-AGI-3 figures, the models tested and the comparisons; opened the official GitHub repository for the MIT licence, the project description and the safety warning. The same figures were found in two independent outlets (MarkTechPost and Crypto Briefing), which agree on 95.5% Best@1, 99.97% Best@3, 183/183 levels and the 8/9, 6/9, 6/9 comparison. Noted the date discrepancy between the announcement (5 August) and the press (6 August). No independent confirmation of the score at the ARC Prize Foundation: the limit is stated in the uncertainties. As per procedure, discarded stories without a primary company source and those already covered by the site.
- Incertezze
- The numbers are vendor claims: at the time of checking there is no independent validation from the ARC Prize Foundation, nor a matching entry on a verified official leaderboard. The margin over the human baseline is one tenth of a point (95.5% against 95.4%), so within the spread of the three runs reported (95.0–95.5). What the RHAE Best@1 metric measures exactly, and the test conditions (number of attempts, compute budget, costs), cannot be reconstructed from outside. The publication date is 5 August according to the official announcement, but part of the trade press reports it as 6 August. The announcement does not say whether the comparisons with proprietary harnesses use the default versions and settings of those tools.
- Perché pubblicarla
- It shifts attention from the model to the scaffolding around it: this is the first time an openly licensed (MIT) agent harness has claimed results on a par with or better than the proprietary tools bundled with the same models, and with fewer tokens. For anyone building agents this is practical news, verifiable line by line, not a closed product announcement; beating the human baseline on ARC-AGI-3 is a strong headline, and that is exactly why it deserves telling with its margins of error in plain sight.