Prime Agent open-source Rust version completes a self-rewrite, with startup speed improved by approximately 13 times
BIBIBI
AT A GLANCE
Prime Intellect released Prime Agent's Rust version, with published tests saying that it starts approximately 13 times faster and uses approximately 83%。
Article
Prime Agent Assemble 2200 agents Agent rewrites itself, starts fast 13 times
Beating AI breaking news, Prime Intellect released the open-source AI coding tool Prime Agent's Rust version. It used 2 weeks to orchestrate 2200 agents,Agent rewriting its own code from TypeScript to Rust. The entire process invoked more than 10000 isolated sandboxes and consumed more than GLM-5.3 through 200000000000 token.
Specifically, a lead orchestrator Agent decomposes tasks and determines their order. Other Agent are separately responsible for planning, coding, reviewing, and verification. The new code is compared with the old version to check the interface display, session records, and model requests. Changes that fail review or testing are sent back for revision. The lead orchestrator Agent does not write product code itself.
Tests published by Prime Intellect show that the time from a cold start to being able to enter a command fell from approximately 738 milliseconds to 56 milliseconds, an improvement of approximately 13 times. Startup memory fell from approximately 607 MB to 106 MB, a reduction of approximately 83%. The new version also adds support for a Windows preview release, and the code is open source on GitHub.
However, automated testing did not find every issue. Over the following weeks, the team continued to have AI fix errors, complete features, and optimize performance, while human engineers identified issues, guided revisions, and reviewed the results.
Original link https://m.theblockbeats.info/flash/371130
Key points
01
Prime Intellect released the open-source AI coding tool Prime Agent's Rust version.
02
The article says that the project used 2 weeks to orchestrate 2200 agents,Agent rewriting the code from TypeScript to Rust。
03
The entire process invoked more than 10000 isolated sandboxes and consumed more than GLM-5.3 through an 200000000000 token。
04
A lead orchestrator Agent is responsible for decomposing tasks and determining their order, while other Agent separately handle planning, coding, review, and verification.
05
Tests published by Prime Intellect show that the time from a cold start to being able to enter a command fell from approximately 738 milliseconds to 56 milliseconds, while startup memory fell from approximately 607 MB to 106 MB。
06
The new version adds support for a Windows preview release, and the code is open source on GitHub.
07
The article says that automated testing did not find every issue; subsequently,AI continued fixing errors, completing features, and optimizing performance, while human engineers identified issues, guided revisions, and reviewed the results.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This was a software rewrite completed through the division of labor among many Agent: the lead orchestrator Agent assigned the work, while other Agent handled planning, coding, review, and verification and compared the new and old versions. The published data indicates that the version switched to Rust starts faster and uses less memory, but automated testing still missed some issues, requiring human engineers to continue identifying and reviewing problems.
Why it matters to readers
This case demonstrates a workflow in which multiple Agent collaborate on code migration, review, and testing; if the published results are reproducible, Prime Agent's startup speed, memory usage, and Windows support have all improved.
Beginners can learn from this that Agent does more than generate code: it can also complete planning, review, and verification through role-based division of labor. At the same time, note that automated testing cannot guarantee that every issue will be found.
Risks and unknowns
The evidence does not provide a specific link to the GitHub repository, so the code and release status cannot be verified on that basis.
The performance figures come from tests published by Prime Intellect; the evidence does not provide the test environment, complete methodology, or independent reproduction results.
Evidence not provided 2200 Multiple Agent, more than 10000 sandboxes and more than 200000000000 token original records.
Related Developments
Loading event timeline…
The scope of functionality and stability of the Windows test version has not been described.
The specific results of subsequent bug fixes, feature completion, and performance optimization are unknown.
Related concepts
Agent
This term is not in the glossary yet. Browse related concepts in the glossary.
