Prime Intellect says its open-source coding agent, Prime Agent, coordinated 2,209 agents to port the tool from TypeScript to Rust and optimize it. The company says the effort involved more than 10,000 Prime Sandboxes and about 228.7 billion tokens through its GLM-5.3 endpoint. A signed GitHub release shows the Rust version, v0.10.0, was published on Oct. 9.
Prime Intellect describes the port as a two-week effort and the separate performance-optimization cycle as three days. Release-quality follow-up continued over the following weeks. The company says its engineers set up the tests that let agents measure whether features still behaved like the TypeScript version. Its team later used the Rust build internally, and engineers stayed involved in directing and reviewing the follow-up work.
According to the company, Prime Agent's root agent divided the rewrite into tasks and merged the results without writing product code. The port involved 1,981 agents and 192.99 billion tokens; the optimization cycle involved 228 agents and 35.70 billion tokens. Across both stages, Prime Intellect reports 16,758 messages between agents.
Each task moved through four roles: a planner specified the feature and its test, an implementer wrote code, a reviewer using a different model looked for errors, and a verifier ran the tests in a fresh sandbox. The checks compared the two versions' terminal screens and model requests, checked that their daemon protocols matched, and audited features that scripted tests could not cover.
Prime Intellect also reports that the Rust version was quicker to start and used less memory in its own benchmark. On a fresh four-core, 8-gigabyte sandbox, the company measured 55.8 milliseconds from a cold launch until typing worked, compared with 737.8 milliseconds for the TypeScript version. Its process-tree memory measurement after startup was 106.0 megabytes for Rust and 607.4 megabytes for TypeScript.
Those measurements came from a custom test in which each agent harness ran in a terminal against a scripted model, so the timings excluded model inference. Prime Intellect says the suite had no common benchmark standard for comparisons with other coding tools. Its speed and memory numbers are therefore company-reported results under that test setup, not an independent ranking of coding agents.
Prime Intellect says internal use later exposed bugs and behavior differences that its parity checks had not exercised. During the follow-up, agents also reviewed beta-user logs to diagnose runtime and agent problems users reported. The company says the Rust release includes native Windows support in beta, and Prime Agent remains open source.
Sources
- Prime Intellect: Rewriting Prime Agent in Rust
- Prime Agent v0.10.0 release
- Image: Rust logo by Looobay, Wikimedia Commons, licensed under CC BY-SA 4.0 (unmodified).