Prime Agent

Every file edit, shell command, tool call and subagent launch runs as code inside a single persistent IPython session, and that session is the only built-in tool the model gets. Prime Intellect calls the pattern a Recursive Language Model: context lives as variables in the REPL instead of being replayed into the prompt, and rlm(...) spawns real child agents whose results come back as ordinary values.

On top sits the Continual Harness, durable state holding supplemental prompts, memories, skill descriptions and subagent specifications. The /refine command reviews the trajectory so far and writes only small, evidence-backed updates into it, keeping snapshots for rollback and leaving the immutable base prompt untouched. Sessions are daemon-backed, so work continues after the terminal detaches and can be reattached later, while heartbeats, schedules, persistent goals and a budget-bounded autonomous mode aim at tasks that outlast one chat window. Skills are importable Python packages, running agents can message one another, and the CLI is MIT-licensed for macOS and Linux. One warning comes from the project itself: model-generated code runs with your permissions, and this is not a security sandbox.