Software & DevOpsDone with Claude Code Oct 7, 2026
Run parallel local coding agent sessions
A developer runs up to 10 coding agent sessions at once, using Claude Code and local models on RTX 3090s with cached contexts for daily development work.
- What They Did
- Roark66 runs coding agents on RTX 3090s with 24GB each, using Claude Code alongside local models. They keep up to 10 agent sessions open at once, though never more than 5 are active at the same moment. Because they control inference, they keep full contexts cached in RAM and can save the cache to disk, so a session isn't re-read in full after sitting idle.
- How It Went
- Holding about 15 full contexts cached in RAM pushes their local cache utilisation to 95% or more. They say their usage of roughly 1.3B input tokens a week would cost thousands of USD at API prices.
- Worth Knowing
- Cloud caches expire after 5 minutes or 1 hour, so an idle 500k context gets billed as full input again; local caching avoids that, but you need enough RAM to hold the contexts.
Try it yourself with Claude Code
I want to work on [N] separate tasks in my repo at [path] at the same time. For each task, create its own git worktree and branch, work on it independently, run the tests, and summarize what changed. The tasks are: [task 1], [task 2], [task 3]. Stop when each branch has a passing diff ready for my review, and do not merge anything.