What they did
Prime Intellect set a single root Prime Agent to run the rewrite with as little human intervention as possible. The human work went into verification: a differential suite that runs the TypeScript and Rust binaries against the same scripted model and diffs the terminal frames, diffs of session transcripts and provider requests, protocol checks, and an agent-led audit of every feature as matching, partial or missing. The root agent wrote no product code; it ordered the tasks topologically and each task went through four agents: a planner, an implementer in its own worktree, an adversarial reviewer on a different model and in a separate context, and a verifier in a fresh sandbox. Failures went back to the implementer. It ran on two 8-core nodes, pushed compilation and diffing out to more than 10,000 Prime Sandboxes, and used about 200 billion tokens from GLM-5.3 on their own inference endpoint. Once parity held, a second orchestrator ran a three-day performance hill-climb with no numeric target.
How it went
The port used about 193 billion tokens across 1,981 agents; the hill-climb another 36 billion across 228 agents, logging 144 experiments and merging 69 improvements. They report time-to-type roughly 13x faster cold and warm, process memory after startup about 5.7x smaller, and a 2.9x smaller install. The parity checks only covered what they exercised, so remaining bugs were found by moving their own agents and then their team onto the Rust build, followed by weeks of follow-up work in which agents still wrote the fixes while humans found the problems and reviewed. The code went from a 15,000-line largest file to nine crates with nothing over 5,000 lines.
Worth knowing
This is a vendor describing its own product rebuilding itself, with a benchmark suite it wrote and flags as non-standard. HN commenters asked for the input, cache-hit and output split of the 200 billion tokens, which the post does not give, so the cost cannot be judged. The planner, implementer, reviewer, verifier loop is what readers asked about most, and the company says it is building that pattern into the product.
GitHub