Prime Agent: An Open-Source Coding Agent That Rewires Context

What Makes Prime Agent Different

Most terminal coding agents work the same way: call a tool, dump the result into context, repeat. That approach burns tokens fast and degrades on long tasks. Prime Agent, built by Prime Intellect, takes a different route it calls a recursive language model, or RLM. The model runs inside a persistent Python session and treats its own context, its tools, and even sub-agents as things it can manipulate programmatically rather than raw text it has to read in full.

Practically, this means the model can write a snippet of Python to process a large file and pull out only what it needs, instead of loading the whole thing into context. Sub-agents work the same way — a parent agent can spawn a full child agent as a function call and get the result back like a return value, enabling parallel or background work without flooding one context window.

• Install it via the one-line curl script on macOS or Linux (no native Windows support; WSL only).
• Authenticate with an existing Claude, ChatGPT, or GitHub Copilot subscription, or use API keys from over 20 providers.
• Go fully local by pointing it at Ollama, LM Studio, or vLLM if you want zero cost.

The Self-Improving Harness

Most agents rely on static, hand-written system prompts and memory files. Prime Agent instead treats its own harness — prompts, memories, skill descriptions, sub-agent specs — as durable state the agent can edit. Running the /refine command reviews recent work and bakes small, evidence-backed corrections into that state, so mistakes you've already fixed once don't recur.

The base system prompt itself stays fixed, and every refinement is snapshotted so you can roll back a bad change. There's also a proper skills system: recurring workflows become executable Python packages rather than markdown notes, with a built-in tool to package them for a project or for personal use.

• Use /refine after correcting the agent to permanently encode that lesson.
• Build skills for repeated workflows instead of re-explaining them each session.
• Roll back any refinement that makes results worse, since changes are snapshotted.

Long-Running and Autonomous Work

Prime Agent is designed to run unattended. A /goal command sets objectives that persist across turns and context compactions, a heartbeat command wakes it periodically, and an autonomous mode lets it work within token and time budgets you define, subject to quality gates you set.

Sessions are daemon-backed, so closing your terminal or losing an SSH connection doesn't kill the run — you can reattach later. Running agents can even message each other directly to coordinate without your involvement.

• Set persistent goals with /goal for multi-session work.
• Reattach to sessions with prime-agent attach after a dropped connection.
• Configure autonomous mode carefully, since it runs with your permissions and has no sandbox.

The Benchmarks and the Caveats

On the ARC-AGI-3 benchmark, Prime Agent running Claude Opus scored 95.5% best-of-one, edging past the reported human expert baseline of 95.4%, and 99.97% with best-of-three. On long-context benchmarks like RULER and LongBench Pro, it reportedly matched or beat Claude Code and Codex, with both proprietary and open-weight models like GLM. It also reportedly built working Sega Genesis and Game Boy Color emulators from spec on an internal emulator benchmark.

These are all figures reported by Prime Intellect itself, not independently verified. The tool is brand new, so rough edges are expected, and it executes model-generated code with your user permissions — there is no security sandbox.

• Treat benchmark numbers as unverified until outside testing confirms them.
• Avoid untrusted repos, or run the agent inside a container or VM.
• Expect rough edges given how new the release is.

The Payoff

If the token-efficiency claims hold up, Prime Agent could mean fewer wasted tokens and less usage-limit anxiety on long coding sessions, at no extra cost if you already pay for Claude or ChatGPT. The bigger win is longer-term: a harness that actually learns your codebase's quirks instead of making you repeat corrections every session.