BuildNot yet confirmed elsewhere1 publisher2 min readPublished
Prime Intellect's 2,000-agent Rust port of Prime Agent is benchmarked only on harness overhead
Prime Intellect says over 2,000 agents rewrote Prime Agent in Rust, using more than 10,000 sandboxes and 200 billion-plus tokens. Its speed and memory gains were measured with inference excluded, so they describe harness overhead and leave coding-task time untested.
The Engineer · Build desk
What happened
- A root agent split the port into dependency-ordered tasks and coordinated separate planning, implementation, review and verification agents, so no agent approved its own changes.
- Parity checks ran the Rust and TypeScript builds through scripted tasks and compared terminal frames, session transcripts, provider requests and daemon-protocol messages.
- On fresh four-core, 8 GB sandboxes, cold-start input-ready latency fell from 736.1 milliseconds for TypeScript to 51.9 milliseconds for Rust.
- Process-tree memory on a 10 MiB session fell from 1,130.0 MB under TypeScript to 237.3 MB under Rust, according to the company's hillclimb table.
Why it matters
- constraint The 684 ms saving comes once per session start, so whether it shortens a customer's coding task depends on the inference time the benchmark left out.
- capability Dropping a large session from about 14 percent to about 3 percent of an 8 GB sandbox leaves room for more persistent sessions and subagents on the same machine.
- decision Teams pricing an agent-driven port have to count weeks of human follow-up alongside tokens and sandboxes, since passing parity tests did not make the Rust build ready to ship.
Both speed figures come from a scripted model, with inference excluded [13]. They measure the harness's startup and runtime overhead. On Prime Intellect's test, Rust's cold start is about 14 times faster than TypeScript's [18]. For that gain to reach a customer, harness startup would have to be a sizable share of a session's wall clock. The test removed the model time that decides that share. RuntimeWire concluded the benchmarks do not establish better end-to-end results for customers [16].
The memory figure has a better chance of transferring. Large-session process-tree RSS fell by a factor of about 4.8 [19]. Prime Agent keeps a persistent daemon that streams model output, runs tools and coordinates agents across sessions. Prime Intellect argues Rust's native execution and compile-time checks suit that long-running workload [15]. We'd expect the saving to count most where many sessions share one sandbox. The rewrite also adds Windows support, the company says [3]. Comparisons with other harnesses are weaker ground. The company's own post says there is no common benchmark standard and those results should be treated cautiously [14].
The process behind the port is the better engineering in the release. The root agent wrote no product code, so it could spend its time on task allocation and integration [6]. Anyone who has worked under a good engineering manager will recognise the job description. Verification used the TypeScript build as the oracle. For a port we think that is the right choice, because the old program is the specification and any divergence shows up as a diff against it [7]. Agents also audited features component by component [8].
The scripts still missed things. Prime Intellect moved its own agents and staff onto the Rust build, found gaps the parity checks had not caught, and spent the following weeks fixing bugs, finishing features and polishing the interface [10]. Engineers directed that follow-up and reviewed the results [9]. The agent run itself took more than two weeks [1].
The scale is large. At the stated floor figures, the run averaged about 100 million tokens per agent [21]. The rewrite's tokens, all from Prime Inference's GLM-5.3 endpoint [2], equal about 2.5 percent of everything Prime Agent has processed since its August launch, by the company's tally [22]. Prime Intellect does not break out active users, retention or the commercial share of that usage [17].
What to watch
- End-to-end task timings with live inference for the Rust and TypeScript builds, from Prime Intellect or an outside tester.
- A shared benchmark for comparing agent harnesses; Prime Intellect says none exists yet.
- Bug reports from Windows users, the platform the Rust build adds.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption30
- Hype gap+15
- Incentives75
- Confidence40
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Prime Intellect says more than 2,000 agents rewrote Prime Agent, its open-source coding agent, in Rust over more than two weeks.
- [2]
The rewrite used more than 10,000 Prime Sandboxes and over 200 billion tokens from Prime Inference's GLM-5.3 endpoint.
- [3]
Published on October 9, the release says the Rust version starts faster, uses less memory and adds Windows support.
- [4]
Prime Agent launched in August and the company says it has since been downloaded more than 300,000 times and processed over 8 trillion tokens.
- [5]
A root agent divided the work into dependency-ordered tasks and coordinated planning, implementation, review and verification agents; separate agents checked the work rather than the same agent writing and approving its own changes.
- [6]
The root agent wrote no product code, so it could focus on task allocation and integration.
- [7]
Prime Intellect compared the Rust and TypeScript versions against scripted tasks, checking rendered terminal frames, session transcripts, requests sent to model providers and daemon-protocol messages.
- [8]
Prime Intellect also had agents audit features component by component.
- [9]
The company says its engineers directed follow-up work, investigated issues found in internal use and reviewed the results; people were not removed from release decisions.
- [10]
Internal testing exposed gaps the automated parity checks had not caught; Prime Intellect moved its own agents and staff onto the Rust version, then spent the following weeks fixing bugs, finishing feature work and polishing the interface.
- [11]
On fresh four-core, 8 GB sandboxes, Prime Intellect's hillclimb table reports cold-start input-ready latency of 51.9 milliseconds for Rust versus 736.1 milliseconds for TypeScript.
- [12]
For large-session process-tree RSS on a 10 MiB session, the table reports 237.3 MB for Rust versus 1,130.0 MB for TypeScript.
- [13]
The measurements use a scripted model and exclude inference, so they speak to the harness's startup and runtime overhead, not how quickly a model completes an end-to-end coding task.
ReportedSupportedSource: RuntimeWire, describing Prime Intellect's benchmark methodView cited source - [14]
Prime Intellect's post says there is no common benchmark standard for agent harnesses and that its cross-product comparison results should be treated cautiously.
- [15]
Prime Agent's persistent daemon streams model output, runs tools and coordinates agents across sessions; Prime Intellect argues Rust's native execution and compile-time checks better suit that workload.
- [16]
Company benchmarks show lower harness startup latency and memory use while excluding inference; they do not establish better end-to-end results for customers.
- [17]
The announcement does not break out active users, retention or how much of Prime Agent's usage was commercial.
- [18]
Rust's cold-start input-ready latency is about 14 times lower than TypeScript's, a saving of about 684 milliseconds.
- [19]
Large-session process-tree RSS is about 4.8 times lower in Rust, a reduction of about 893 MB (about 79 percent).
- [20]
On an 8 GB sandbox, the TypeScript harness's large session occupies about 14 percent of RAM and the Rust one about 3 percent.
- [21]
At the stated floor figures, the rewrite averaged about 100 million tokens per agent.
- [22]
At the stated floor figures, the rewrite's tokens equal about 2.5 percent of the tokens Prime Agent has processed since launch.
Sources
1 independent publisher whose own reporting we read for this story.
- runtimewire.comPrime Intellect says 2,000 agents rewrote its coding agent in Rust
1 article · October 9, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.