Prime Intellect open sources Rust-built Prime Agent for long-running coding work

Prime Intellect has released Prime Agent, an open source coding and research agent built for work that stretches over long periods rather than a single prompt-and-response cycle. According to a post on dev.to, the project has been rebuilt in Rust, and its current stable release, v0.10.2, dates from 28 October 2026. The repository has accumulated more than 22,000 stars, and the README lists the project under an MIT licence.

A reviewer that is not the writer

The idea Prime Intellect places at the centre of the design is that the component producing code should be different from the one evaluating it. Instead of a single loop that generates a change and then inspects its own work, Prime Agent assigns writing and review to distinct roles. The dev.to post presents this as the project's founding rule, and it advises anyone trying the tool to pick a genuinely long-running task first and to keep the review step apart from the writing step.

The underlying motivation is familiar to anyone who has watched an agent spiral: a model reviewing its own output faces a built-in conflict of interest. Splitting the roles gives the reviewer a narrower job and removes the temptation to rubber-stamp the writer's work. The announcement does not include benchmark numbers comparing the split design against single-loop agents, so the claimed benefit rests on the architecture for now.

Context as variables, subagents as functions

Mechanically, Prime Agent is organised around what Prime Intellect calls the Recursive Language Model. Context behaves like variables, and subagents are called the way a program calls functions, all within a persistent Python REPL. Rather than one ever-growing prompt, the main agent hands scoped pieces of work to subagents and receives their results back, much as a function returns a value.

That model pairs naturally with long sessions. Work that runs for hours cannot depend on a single context window, so breaking the task into callable units keeps each piece small while the persistent REPL holds the shared environment together.

State that survives between sessions

The second half of the architecture is what the project calls the Continual Harness. Memories, skills, prompts and subagent definitions are stored as durable state, so an agent returning to a task later does not start from scratch. According to dev.to, this state is meant to be refined incrementally, through small updates supported by evidence rather than wholesale rewrites of what the agent has learned.

An agent rewrite performed by agents

One detail in the announcement stands out: Prime Intellect says the Rust rewrite itself was carried out by more than 2,000 agents. That figure is the company's own claim rather than an independently verified count, but it is a striking demonstration of the coordinated, extended work the tool is meant to support, in effect the project rebuilding itself with the workflow it advertises.

Why it matters

Two chronic weaknesses in today's coding agents are context that degrades as a session grows, and a single actor that both produces and grades its output. Prime Agent takes a concrete swing at both: the Recursive Language Model keeps each context small by delegating to callable subagents, and the writer-judge split turns review into an independent step rather than a formality.

For teams with tasks that legitimately run long, such as migrations, large refactors or extended research sweeps, the design is worth a look, and the MIT licence plus the open repository make it cheap to inspect. The separation of writer and reviewer is the most portable idea here: it costs little for other agent frameworks to adopt, and it targets a failure mode almost everyone who has used coding agents has seen firsthand.