Prime Intellect ships Prime Agent, an open-source coding harness that rewrites itself

Prime Intellect has released Prime Agent, an open-source coding harness built on the argument that today’s agent scaffolding was designed for weaker models. Fixed tool-calling schemas and context compaction, the company says, make frontier models fight their own wrappers. Prime Agent’s answer rests on two ideas. The first, a Recursive Language Model (RLM), gives the model a persistent IPython REPL as its single tool: sub-agents, skills, and history all become variables and function calls the model manipulates in code rather than through rigid schemas. Because context lives in the kernel, the agent can keep working across very long sessions without losing access to its earlier state. The second, the Continual Harness, lets the agent create, read, update, and delete its own prompts, skills, memory, and sub-agents mid-run — so the scaffolding adapts to what the agent learns instead of being frozen at design time.

Under the hood, a background daemon owns every live session over a local socket, and each session tree runs in a recoverable worker that can be rebuilt from append-only JSONL logs and a kernel snapshot if it crashes. Sub-agents are themselves full Prime Agent instances, spawned asynchronously via an rlm() call that returns a handle immediately; agents then coordinate through message passing, including agent-to-agent communication across separate sessions (restricted to a parent–child “nuclear family” to prevent cross-talk). A TUI with an Agents View lets users attach, detach, and steer any running, idle, or unloaded session, with inactive sub-agents paged out of memory and reloaded on demand.

The broader bet is that harness design should anticipate model capability rather than trail it. By exposing sub-agent orchestration as ordinary async code and letting the agent restructure its own memory and tooling, Prime Intellect is positioning Prime Agent both as a practical coding assistant and as a substrate for long-horizon autonomous work and research — with the expectation that future models trained around these primitives will need far less hand-holding.