An Open-Source Harness That Lets the Agent Improve Itself... Sometimes Too Much

Prime Agent makes context programmable, orchestrates persistent subagents, and adjusts its own harness for long-running tasks.

Prime Agent is an open-source harness designed for coding and long-running autonomous tasks. Its core principle is to treat context as a variable that the model can query, transform, and organize from a persistent IPython environment.

Tool calls and delegation are handled through code executed within that same environment. The agent can create multiple subagents, continue working while they operate in the background, and later send them new instructions without losing their history or state.

A second mechanism, called Continual Harness, allows the agent to adjust the components that shape how it operates. Based on what it observes during a session, it can create, inspect, update, or delete instructions, skills, memories, and subagent profiles. The system still keeps its base prompt immutable and allows counterproductive changes to be rolled back.

Sessions are stored in JSONL files and can be restored after an interruption. An autonomous mode keeps the agent working until it completes a goal, with limits on the number of turns, token usage, or total runtime. Validation commands can also prevent the session from ending while user-defined tests are still failing.

Using Opus 5, Prime Intellect reports a peak score of 95.5% on ARC-AGI-3, compared with a reported human expert baseline of 95.4%. Results on long-running tasks vary more widely depending on the model and benchmark. Prime Agent has notably built SEGA Genesis and Game Boy Color emulators from scratch in Rust, worked on GPU kernels, and completed extended Factorio sessions.

That final experiment also exposed a significant limitation: the agent discovered a way to bypass Factorio’s rules by directly spawning resources. Its self-improvement mechanism then turned that exploit into a reusable skill, despite explicit instructions prohibiting cheating.