Prime Agent

What it is

Prime Agent is Prime Intellect's open-source harness, released in August 2026 alongside the company's Recursive Language Model paper (arXiv:2608.23552). It is the reference implementation of that paper rather than a general-purpose product, and it shows: the design is argumentative, not accommodating.

The argument is that a harness should stop handing the model a toolbelt. Most coding agents expose a dozen or so functions — read a file, write a file, run a command, search the repository. Prime Agent exposes one: a persistent IPython kernel. There are no filesystem tools by default. If the model wants to read a file, it writes Python that reads a file.

How the loop works

Each turn the model emits Python, the kernel runs it, and the output comes back. Because the kernel persists across turns, so does everything in it — variables, imports, loaded data, open connections. The model is not re-establishing context on every step; it is working in a session it built earlier. A Python package, prime-agent-runtime, lives inside the kernel and gives the model the harness-side helpers it needs.

Two things sit on top of that. The first is the continual part: the agent rewrites its own prompts, memory and skills while it runs, through a refine step that stores versioned entries and can roll them back. The operating instructions are treated as mutable state, not as a fixed header. The second is recursion: the agent can spawn child agents as background sessions, each with its own model and its own session directory, and collect their results.

Three integration modes are provided — ACP for editor and client integration, RPC for programmatic control, and a JSON mode for embedding the loop in something else.

Isolation and permissions

There is no meaningful sandbox here, and it would be wrong to read the kernel as one. A persistent IPython process has exactly the rights of the user who started it, and arbitrary Python is a superset of anything a permission prompt could gate. Child agents get their own session directories, which keeps their working state apart, but they run on the same host with the same reach. Treat the container or VM you launch it in as the actual boundary.

Who it is for

People running long unattended jobs who are comfortable with Python as the interface layer, and researchers interested in harnesses that modify themselves. If you already drive agents by writing scripts for them, this removes the translation layer.

Limits

The headline numbers are author-reported. Prime Intellect report 95.5% Best@1 on ARC-AGI-3 RHAE against 30% for the previous best, and they report favourable comparisons against popular harnesses. We have not seen an independent reproduction of either, and until one exists these should be read as claims from the team that built the thing, not as results.

Beyond the numbers: a harness that rewrites its own prompts is hard to reproduce runs on, and hard to debug when a run goes wrong — the version history helps, but you are still debugging a moving target. The project dates from August 2026, so the usual new-project caveats apply.

Alternatives on this board

• OpenCode — provider-agnostic with a conventional tool set, if the single-kernel idea is too much.
• Claude Code — the permissioned-tools approach Prime Agent is arguing against.
• DeepSeek Harness — the other 2026 harness built around a strong architectural opinion.

Sources

• Recursive Language Model paper (arXiv:2608.23552)
• Prime Intellect — the repository is linked from the paper and the project site