Overnight Compute

overnight-compute

SQLite-backed compute leases for running multiple coding agents on one shared machine overnight.

I made this because I was tired of telling a pile of agents "you get the GPU for the next two hours, then you get it after that" and hoping they would not step on each other. `overnight-compute` gives them a tiny shared queue, lease, heartbeat, and early-handoff protocol.

Install

curl -LsSf https://astral.sh/uv/install.sh | sh
mkdir -p ~/.local/bin
curl -fsSL https://raw.githubusercontent.com/Infatoshi/overnight-compute/main/bin/overnight-compute -o ~/.local/bin/overnight-compute
chmod +x ~/.local/bin/overnight-compute
overnight-compute init

Make sure `~/.local/bin` is on `PATH`.

The CLI is a standalone Python script launched through `uv`.

Quick Start

Schedule a 12-hour queue with six 2-hour slots:

overnight-compute schedule-chain --start now --slot 2h \
  pretrain-a pretrain-b pretrain-c pretrain-d pretrain-e pretrain-f

Each agent should wrap expensive commands:

overnight-compute run --agent pretrain-a -- uv run pytest

For manual multi-step work:

nvidia-smi
overnight-compute wait --agent pretrain-a
# do expensive work
overnight-compute heartbeat --agent pretrain-a --ttl 30m
overnight-compute release --agent pretrain-a --status done

Check the queue:

overnight-compute status

Pasteable Agent Prompt

I am going to sleep for 12 hours. You need to operate autonomously, coordinate compute with other agents, and make useful progress without asking me questions unless you are truly blocked by something only I can resolve.

First, create your own overnight-compute agent name. Make it specific to the experiment or workstream you are actually going to run, not generic. Use lowercase letters, numbers, and hyphens only. Keep it short but descriptive.

Before any GPU work or long CPU work:
1. Run `nvidia-smi`.
2. Check the queue with `overnight-compute status`.
3. If you ar