Prime Agent

• Running an overnight autonomous refactor with `--autonomous`, a turn budget, and `npm run check` as the gate that must pass before it stops
• Fanning out subagents to summarize different parts of a large codebase in parallel and collecting their findings in the parent session
• Running long-horizon research evaluations or benchmarks with goals, heartbeats, and budgets, then inspecting every session in the Agents View
• Letting `/refine` turn a recurring fix, such as retrying a flaky test, into a reusable memory or skill for later sessions
• Processing long logs or documents programmatically from variables in the REPL instead of reading them into the context window
• Leaving a long task running in the background daemon, closing the terminal, and reattaching later with `prime-agent attach`

Prime Agent is one of the most research-driven open-source terminal agents: rather than adding tools, it gives the model a Python REPL and lets it orchestrate subagents, process context, and edit its own harness state in code. That design pays off on long, autonomous, or context-heavy work, where background sessions, budgets, gate commands, and parallel subagents matter more than a polished interactive experience, and it makes Prime Agent a natural harness for evaluation work. For everyday pair programming, the benefit is less clear, since no model has been trained around this design and the project ships changes almost daily. Treat the self-improvement as a useful, auditable memory and skills layer rather than a guarantee of better results, run it in a sandbox or disposable worktree, and read its published benchmarks as Prime Intellect's own measurements.

About Prime Agent

Prime Agent is built on top of the Pi coding agent and organized around two ideas from Prime Intellect's research. The first is the Recursive Language Model (RLM): the model's only built-in tool is a persistent IPython kernel, its context and session history are available there as variables, and file edits, shell commands, MCP tools, and subagents are Python functions. Calling `rlm.spawn(...)` starts a real child agent with its own model, kernel, and history, so the agent can fan work out in parallel, keep children around for follow-ups, and message parents, siblings, or children directly. The second is the Continual Harness: supplemental prompt notes, memories, skills, and subagent specifications are stored as durable state the agent can create, update, and delete. `/refine` reviews the session and applies a small, recorded edit to that layer, never to the immutable base system prompt, and a bad refinement can be rolled back by ID. Around this sits infrastructure for long-running work. A background daemon owns every session, so agents keep running when the terminal closes and can be reattached. An Agents View lists running, idle, and saved sessions and their subagents. `/goal` keeps an objective active across turns, heartbeats and schedules re-enter a session later, and autonomous mode keeps working within turn, token, and time budgets and can require a command such as a test suite to pass before it stops. You sign in to model providers with `/login`; the repository lists clients for Anthropic, OpenAI, Google, Amazon Bedrock, Mistral, Azure, and Prime Intellect's own Prime Inference, and release notes add ChatGPT (Codex) and xAI Grok subscription sign-in. Prime Intellect also positions it as a harness for research evaluation. Its launch post reports 95.5% RHAE Best@1 on ARC-AGI-3 with Opus 5 in autonomous mode, the best of three runs (95.0, 95.2, 95.5) with only the task prompt adapted, and the arXiv paper describes the harness as raising ARC-AGI-3 RHAE Best@1 from 30% to 95.5%. These are Prime Intellect's own results. The tradeoffs: it runs model-generated Python and commands with your user permissions, and the README states its worker and kernel processes are not a security sandbox. It is pre-1.0 and changes almost daily. The one-line installer currently ships the TypeScript release, while the repository's main branch holds a Rust port that is published only as nightly builds. Prime Intellect also notes that no model has yet been trained around this harness design, so it may be less predictable with some models than their native harnesses.