Prime Intellect open-sources Prime Agent, a self-refining RLM

Prime Intellect has published Prime Agent as an open-source, MIT-licensed coding and research agent, and the project's own README sums it up as "a self-improving RLM agent for coding workflows and long-running autonomous tasks." The interesting bit is not the tagline, because every coding agent claims some flavour of that, but how the repo defines "self-improving" and "RLM," because both terms are load-bearing.

The RLM (Recursive Language Model) framing treats the agent's context as variables inside a persistent Python REPL, with subagents callable as functions via `rlm(...)`. Skills ship as importable Python packages rather than YAML or markdown recipes, so recurring workflows compile down into ordinary Python you can inspect and diff. In practice that makes the agent's memory feel more like a live notebook than a chat transcript.

"Self-improving" is the phrase doing the most marketing work, and Prime Intellect's version of it is narrower than it sounds. A `/refine` command reviews the agent's own trajectories and, per the README, "applies updates to harness state without rewriting the base system prompt." The updates land as supplemental prompts, memories, skill descriptions, or subagent specs, with a recorded refinement history you can inspect. That is a more auditable claim than "the model retrains itself," and it is the right claim to make.

The README does not publish benchmark numbers, and it does not name which base LLMs the agent is designed against, so the self-improving pitch is not yet backed by third-party evidence you can point at. Daemon-backed sessions with autonomous mode and persistent tool access also open real new attack surface, and anyone running this on shared infrastructure will want the "bounded autonomous mode" budgets configured tightly before leaving it running overnight.

Two of the researchers in our Who's Who directory circulated the repo already, which is roughly the level of attention any credible open agent release now attracts on day one. If the /refine loop turns out to durably improve task success rather than just accumulate stale lessons, an MIT-licensed local runtime becomes a genuine alternative to the closed hosted agents, and that is the outcome to watch for.