Prime Agent

About

Prime Agent exists because long tasks break harnesses that hold all context in a shrinking window: it instead puts the model in a persistent IPython REPL where the prompt is a variable, tools are functions, and rlm(...) spawns child agents that run in parallel and message each other directly without routing through the user. Everything is programmatic — file operations, shell commands, context management — which makes the agent's behavior inspectable and composable rather than a sequence of opaque tool calls. A Continual Harness accumulates what works across a session, and /refine reviews the trajectory and applies small evidence-backed updates to memories, skills, and subagent specs, with snapshots for rollback and the base prompt left untouched. Sessions are daemon-backed with heartbeats, schedules, persistent goals, and a bounded autonomous mode with quality gates. Prime Intellect built it for long-running research and evaluation work alongside its Verifiers and PRIME-RL stack, and the README warns plainly that it is not a security sandbox.