Prime Intellect launches Prime Agent, a self-improving coding harness built on recursive sub-agents

Prime Intellect today introduced Prime Agent, an open-source coding harness designed around two abstractions: the Recursive Language Model (RLM) and the Continual Harness. The core idea is that modern scaffolding should adapt to what the model learns while running, rather than being frozen at design time. Instead of a fixed set of tools and hand-engineered sub-agents, Prime Agent gives the model a persistent IPython kernel as its only tool, with everything else—sub-agents, skills, memory, even prompt notes—exposed as callable functions inside that kernel.

The result is a harness that treats the agent's own scaffolding as a living system. Sub-agents are spawned with a simple await rlm("sub-task") call, each running as a full Prime Agent instance with its own kernel, session tree, and history. These sub-agents can persist beyond their initial call, letting the parent agent message them later by session ID. The background daemon manages all sessions over a local socket, supports detaching and reattaching, and recovers crashed workers from JSONL logs and kernel state snapshots.

Self-improvement through /refine

The most distinctive feature is /refine, a pipeline that reads the agent's own trajectory and applies the smallest relevant update to improve future outcomes. It can update a prompt note, memory, skill, or sub-agent spec, and each refinement records its trigger and result, so changes are evidence-backed. This moves beyond static, hand-tuned agent definitions toward a harness that evolves with use.

Prime Agent is available now and can be installed via the project's GitHub. It's designed to work with both open and closed frontier models, and the team expects the feature set to become more powerful as future models are trained around it.

For developers, the shift to programmatic control is significant. Instead of designing rigid tool schemas that constrain how a model can act, Prime Agent lets the model write code to orchestrate its own workflows—parallel fan-out, background tasks, and persistent sub-agent communication become part of the model's natural toolkit. That's a meaningful step toward agents that don't just execute a plan, but improve the plan as they go.

Why it matters

The harness is a bet on model capabilities continuing to outpace the scaffolding around them. If frontier models can handle open-ended programmatic control, then rigid tool-calling schemas are an unnecessary bottleneck. Prime Agent's approach—treating the harness as mutable state that the agent itself can edit—is a direct response to that. It's early, but it points toward a future where the agent's own environment is something it can debug and optimize, not just work within.