Prime Intellect ships a Rust rewrite of Prime Agent, claiming 13-fold faster startup
The company’s large agent swarm handled the port, but scripted checks did not catch every problem. Human-guided follow-up work brought the release into shape.
Loading page…
The company’s large agent swarm handled the port, but scripted checks did not catch every problem. Human-guided follow-up work brought the release into shape.
Listen to this story
Prime Intellect released the Rust-based Prime Agent on October 9, 2026, with a narrowly scoped performance claim: its custom benchmark found keyboard input became available about 13 times faster, not that coding tasks finished faster. The two-week rewrite involved more than 2,000 agents, 10,000-plus sandboxes and over 200 billion tokens, but reaching release quality still took weeks of human-directed fixes beyond the behaviors covered by tests. The release includes isolated, persistent session workers; reusable developer-defined workflows remain a promised next step.
Each task moved through planning, implementation, independent review and fresh-sandbox verification; failures sent work back for revision.
A three-day post-port optimization loop logged more than 144 experiments and audits, merging over 69 changes judged valid.
The release uses nine Rust packages, isolated session workers and on-disk persistence; native Windows support remains in beta.
Prime Intellect shipped a Rust rewrite of Prime Agent on October 9, 2026, turning a large agent-driven migration into a released coding tool. The company says more than 2,000 agents handled the two-week rewrite. Its own benchmarks show roughly 13-fold faster access to keyboard input, though bringing the software to release quality required further human-guided work.
Prime Agent is the software around a coding model that coordinates sessions, tool calls and other agents. Prime Intellect says TypeScript helped it ship quickly, but CPU-heavy rendering and session parsing competed with keyboard input. Each process also carried the overhead of a JavaScript runtime and its memory-cleanup system.
The company chose Rust for native-code performance and checks that catch certain mistakes before software runs. Rust’s compiler checks how data moves between threads and which parts can share it. Prime Intellect says those guarantees matter especially when agents write most of the code, rather than people inspecting every change.
The migration ran across more than 10,000 Prime Sandboxes, using over 200 billion tokens through the company’s GLM-5.3 inference endpoint. A root agent divided the work into tasks ordered by their dependencies. It wrote no product code itself, instead tracking work, merging completed changes and working with humans on priorities and decisions.
Each task passed through four roles. A planner specified the feature, the original TypeScript behavior and its checks. An implementer wrote Rust code in a separate working copy. A reviewer challenged the proposed change using a different model and separate context. Finally, a verifier compiled it and ran tests in a fresh sandbox.
A failed review or verification sent the task back to the implementer; both had to pass before a change merged. Tests ran the old and new versions against the same scripted model, comparing terminal screens, session transcripts and requests to model providers. Other checks covered messages between components, while agents audited features for matches, partial implementations and omissions.
Those checks reduced human review, but they could only verify behavior they exercised. Moving the rewrite agents, then the internal team, onto Rust exposed bugs and missing behavior outside the tests. Bringing the rewrite to release quality took follow-up work over subsequent weeks, with humans finding problems, directing changes and reviewing results.
Once the versions matched on the required checks, a second orchestrator ran a three-day optimization loop. Agents profiled bottlenecks, tested candidate changes against the current version and checked nearby benchmarks for regressions. Prime Intellect says the loop recorded more than 144 experiments and audits, merging over 69 valid changes.
Most gains came from moving work away from startup and rendering, replacing repeated polling with waits triggered by events, and freeing memory after large sessions loaded. Tests used fresh four-core, 8 GB sandboxes and scripted model responses, excluding inference time. Prime cautions that comparisons against other coding tools lack a common benchmark standard.
The release also changes the structure: nine Rust packages have enforced one-way dependencies, and shared message definitions keep components aligned. Each session runs in its own worker process, so one session’s failure leaves others running. Sessions persist on disk, allowing clients to reconnect after a restart. Native Windows support is in beta, and Homebrew installation is available.
Prime Agent remains open source, with installation instructions in the release post. The next promised step is making the orchestration pattern reusable inside Prime Agent: workflows that developers can define once, reuse and update. That packaging remains a plan, distinct from the Rust tool shipping now.
Loading discussion...
Join the conversation
Explain where you would draw the line between autonomy and assistance.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.