Prime Intellect rewrote its coding agent in Rust, 14x faster cold start

A coding agent that takes most of a second to start is noticeable every time the user calls it. Prime Intellect rewrote its Prime Agent harness from TypeScript into Rust to close that gap, and used Prime Agent itself to do the work.

The post, by Kevin Thomas, Sebastian Müller and Seth Karten of Prime Intellect, says the new Rust version is open source under the same repository on GitHub, which carries 21,800 stars, and lists installation for Linux, macOS and a Windows beta. The company says Prime Agent has been downloaded more than 300,000 times and has processed over 8 trillion tokens since launch.

What the numbers say

Prime Intellect reports five before and after figures from the move to Rust, all presented as its own measurements:

• Cold start: 736.1 ms to 51.9 ms, which the post calls a 14.18x speed-up.
• Warm start: 552.3 ms to 41.4 ms, called a 13.34x speed-up.
• Memory on a large session: 1,130 MB to 237.3 MB, a 4.76x reduction.
• Install size: 172.1 MB to 59.6 MB, a 2.89x reduction.
• Agents view screen response: 78.0 ms to 12.8 ms, a 6.09x speed-up.

Each figure is Prime Intellect's own and comes from its own benchmarks. The post does not describe the hardware, the version of Node compared against, or how many runs the medians come from.

How the rewrite was done

The team used Prime Agent to port Prime Agent. The post says more than 2,000 agent instances ran in parallel across more than 10,000 Prime Sandboxes, drawing on Prime Inference's GLM-5.3 endpoint, and used more than 200 billion tokens over two weeks for the core rewrite. More performance work followed over several weeks.

The result is also smaller to read. The longest source file in the Rust version is about 2,500 lines, down from about 15,000 in the TypeScript version. No file exceeds 5,000 lines; the TypeScript codebase had four that did.

For a team already using Prime Agent, there is nothing to decide: the current install from the website now gets the Rust build. For a team shopping for a coding agent harness, the useful question is whether the measured gains survive a second party running them on their own workload. The post includes no third-party reproduction and does not propose a method for one.

Source

• Prime Intellect, Kevin Thomas, Sebastian Müller and Seth Karten, Rewriting Prime Agent in Rust
• GitHub, PrimeIntellect-ai/prime-agent