041 · TOOLS FOR AGENTIC RESEARCH

Tools

Our research runs on agents. These are the tools we built so they can do it end to end, each one a command line first.

01 · Write

SexpGPU

The ML framework built for agents.

Your agent writes the whole experiment as one Lisp file. The compiler checks every shape in milliseconds on a laptop, so mistakes die before they reach a GPU. Then a Rust runtime trains it faster than torch.compile.

$ curl -fsSL https://sx.041.io/install.sh | sh$ sexpgpu check minimal.sx$ sexpgpu run minimal.sx

sx.041.io llms.txt

02 · Run

rrun

Your command on the cheapest GPU that fits.

Say what hardware you need. rrun finds the cheapest machine on the clouds you connect, streams the logs back and deletes the machine when the work is done.

$ curl -fsSL https://rrun.041.io/install.sh | sh$ o41-rrun run --gpu RTX5090,A6000 --spot -- python train.py

rrun.041.io llms.txt

03 · Keep

Artifacts

Train where GPUs are cheapest. Resume anywhere.

Buckets you own in AWS, GCP, Azure or any S3 endpoint become one checkpoint store. Every write lands next to the GPUs, so a preempted run resumes in any cloud with one read, not an egress bill.

$ npx @o41/artifacts versions experiments/run-1234$ npx @o41/artifacts get o41://experiments/run-1234/latest ./ckpt

artifacts.041.io llms.txt

04 · Read

Metrics

Experiment tracking your agents can use.

Log from Python in one line without ever blocking training. Read it back as JSON from the command line, as SQL in DuckDB or from Python, and share the exact charts in the web app.

$ o41-metrics experiment list --search lr3e-4$ o41-metrics series render mamba-lr3e-4-seed7 --name loss -o loss.png

metrics.041.io llms.txt

Built for agents

Hand your agent the URL. It reads the rest.

Every tool answers in JSON, and every docs page answers in markdown when an agent curls it. Each site lists its pages in /llms.txt, so an agent reads only what it needs.