Prime Intellect released a Rust rewrite of its coding agent

TOP 3

What shipped: Prime Intellect’s October 9 account describes a Rust rewrite of its open-source coding agent; the v0.10.0 release was published October 10 in Seoul. The new architecture runs each session in a worker process under a supervisor, uses a shared protocol definition across components, and adds beta Windows support.
- How the rewrite was checked: Prime Intellect says an orchestrator assigned planning, implementation, adversarial review and independent verification to separate agents. Differential tests compared the Rust and TypeScript terminal frames, model requests and session transcripts; the company reports 2,209 agents across the rewrite and performance work. Those checks covered scripted flows, not every user behavior: the team says internal use later revealed gaps, and humans still directed priorities and reviewed results. The published startup and memory comparisons use Prime Intellect’s own runtime suite rather than a common independent benchmark.

GPU scheduling: In an October 9 technical account, Ai2 says it replaced priority-based scheduling with manager-assigned GPU-time budgets, hierarchical fair-share allocation and protected minimum runtimes after which jobs can be preempted. In its reported 30-day test, teams received 98% of owed GPU hours while cluster occupancy remained 98%; the observed debug-job queue improvement had a small baseline sample. Interactive sessions became harder to preserve under preemption, and Ai2 says it is investigating possible fragmentation for large jobs.
Agent handoffs: The SpecWeave v3.0.7 release, published October 10 in Seoul, adds a command to configure Claude Code’s autocompaction threshold and changes its auto-handoff hook to trigger again at a usage threshold. The maintainer says the configuration targets long-context sessions; this is a project release note, not a measured saving for every user.