Agent farms, without the meter.
An orchestrator hands out tasks, dozens of agents work in their own worktrees, and whatever passes the tests gets merged. The kind of workload per-token bills punish most.
Pre-launch. Accounts are open; the API and checkout open soon.
How it runs.
01
An orchestrator splits the work.
A strong model breaks a big change into tasks, each with a test that says when it’s done.
02
Agents take a task each.
Dozens work in parallel in their own worktrees, with quick subagents to search and run tests.
03
CI decides.
What passes goes up for review; what fails goes back in the queue.
Why lanes fit it.
Parallel is the point.
Lanes are sized by how much runs at once, which is the number that matters when you run a farm.
The heaviest workload, flat.
A farm goes through tokens around the clock. On lanes the bill is the plan, however many tokens that is.
A model for each job.
The strongest model plans, quick ones do the legwork, and each takes its own number of lanes.
Works with what you use.
OpenAI Python
import osfrom openai import OpenAI client = OpenAI( base_url="https://api.voidstone.net/v1", api_key=os.environ["VOIDSTONE_API_KEY"],) stream = client.chat.completions.create( model="kimi-k3", messages=[{"role": "user", "content": "Explain this stack trace."}], stream=True,)for chunk in stream: print(chunk.choices[0].delta.content or "", end="")