Prime Agent: A self-improving RLM agent

Most coding agents you've used are basically frozen in amber. Someone designs the prompts, the tool schemas, the memory system, once, and the model has to live inside that cage no matter how much smarter it gets. Prime Intellect's new release, Prime Agent, tries to break that pattern. It's a fully open-source harness built around two ideas the team calls the Recursive Language Model and the Continual Harness, and the pitch is that the scaffolding itself should keep learning, not just the model bolted onto it.

The first piece treats the model's own context as something it can manipulate directly. Instead of calling tools through some rigid API, Prime Agent gives the model a persistent IPython kernel as its only real tool, and everything else, sub-agents included, gets invoked as ordinary function calls inside that REPL. Spawn a sub-agent with something like await rlm('summarize the auth flow'), and it comes back immediately with a handle rather than an answer; the actual result arrives later through an agent-to-agent message. That lets a model fan out several specialist sub-agents in parallel, steer one mid-task, or park a child session for later without losing it, since sub-agents persist with their own kernel and history even after their first task wraps up.

The second piece, Continual Harness, is where the self-improvement claim actually lives. Prime Agent exposes its own prompts, sub-agents, skills, and memory through a create-read-update-delete interface, callable from inside the same kernel the model already uses. A built-in pipeline called /refine watches the agent's trajectory and, when it spots a repeated failure or a useful trick, makes the smallest edit that fixes it, promoting a retry pattern into a reusable skill, say, rather than rewriting the whole system prompt. Every edit gets logged with its trigger and outcome, and can be rolled back by ID if it turns out to be a bad call. The base system prompt stays untouched; only the layer around it evolves.

Underneath all this sits a background daemon that owns every live session over a local socket, so you can detach from a chat without killing the agent loop, and an Agents View, opened with a left-arrow tap, that lets you drill recursively into any sub-agent's own sub-agents. Idle sessions get unloaded from memory after thirty minutes and reload instantly the moment anyone pings them, which matters once you've got nested swarms of agents running. Cross-session messaging is deliberately restricted to what the team calls the nuclear family, parents, children, siblings, to stop unrelated sessions from talking to each other unsupervised.

For unattended runs, Prime Agent adds an autonomous mode reachable straight from the CLI: set a goal, a gate command like npm run check, and a turn cap such as --autonomous-max-turns 20, and the agent keeps working until it explicitly declares the goal complete or hits its budget. Prime Intellect is testing this setup against ARC-AGI 3, a benchmark built around learning the rules of simulated worlds on the fly, comparing Prime Agent's autonomous loop to each model's native harness. The team is upfront that no current frontier model was actually trained with this harness in mind, so today's numbers are really a baseline for what comes once one is.