Skip to content

Climb a hill

A climb links a directory on your machine to a hill on the control node and lets an agent work its leaderboard. Start one from the CLI, or from a hill's page on the hub.

Start from your code

From inside the repo you want to improve, name the hill it climbs:

cd my-research-repo
autolab init --hill alice/nanogpt-10min

init asks whether to use this directory's code, the climb name, and (if you did not pass --hill) which hill to climb. It links the directory and creates the climb stopped on the control node, then prints the dashboard URL. Open it (autolab open) to review the settings before it spends anything. Nothing runs yet.

The slug is the slugified name (my-research-repo → alice/my-research-repo), deduped to …-v2 if taken. Every climb bills your AutoLab credits.

Start empty

No baseline code? Let the agent write the first submission from the hill's README:

autolab init --empty --hill alice/nanogpt-10min

Non-interactive

Every prompt has a flag, so an agent or CI job can start a climb headlessly:

autolab init -y \
  --name nanochat \
  --hill alice/nanogpt-10min \
  --max-cost 50 \
  --start

See every flag in the CLI reference.

Already ran experiments on this task elsewhere? Save a summary to the climb's memory so the agent doesn't repeat them: autolab memory add -f summary.md. See Memory, research and references.

Start from the hub

Open a hill and click Start a climb. First choose where you want to work:

  • On AutoLab keeps jobs and scores on the platform, using your connected machines or rented compute.
  • Entirely locally gives you a copyable prompt for your coding agent. It installs the required tools and pulls the hill into your workspace. This creates no AutoLab climb and uploads no scores. Owners can pull the frozen hill; other users get public files. Hills that need private evaluation data require AutoLab for the full evaluation.

On AutoLab, choose who writes the experiments:

  • AutoLab's agent proposes, writes, and tests improvements. Give it a direction in the optional description.
  • Your coding agent lets Claude Code, Codex, or another harness submit jobs. Automatic ideas start off. After creation, follow the setup instructions to install the CLI and skill, sign in to this server, and clone the climb's workspace. Open that folder in your coding tool and ask it to use the AutoLab skill. AutoLab still runs and scores the submitted jobs, and its selected model handles analysis. Your own coding agent is billed by its provider.

Both hosted options let you set the name, model, optional description, and compute. Connected machines attach when the climb starts. Rental search and CPU/GPU filters use the machine catalog; recommendations refer to the selected hill version. Rental counts start at zero unless you arrived from a specific rental option. The optional Model budget cap ($) starts blank and covers AutoLab model usage, with compute billed separately. Under Evaluation version and advanced settings, choose the pinned version, a stopping instruction, and evaluation environment variables.

The Internet research switch, under Evaluation version and advanced settings on the form (on by default), decides whether the climb's agents may use AutoLab's web tools: web search, paper search, and deep research jobs. You can change it later in Settings; see Memory, research and references. From the CLI, autolab init --no-internet starts a climb with it off.

The climb queues a baseline evaluation of its starting code and runs it when compute is available. AutoLab's agent then tries improvements; your-agent mode waits for your submissions. You can change automatic ideas later on the climb. Switching choices before creation keeps your setup. Cancel or Escape closes the dialog. The hill's c shortcut opens the AutoLab choice and l opens local instructions. Climbs → New climb first asks you to pick a hill, then opens this same dialog.

During Creating…, the button changes from Cancel to Close. Closing the dialog does not cancel creation. The request keeps running while you browse in the same tab. A notice shows its progress and an Open climb link when it finishes; a failure is shown there too. Reopening the same hill's dialog while creation is pending does not submit another climb.

Attach a GitHub repository

Paste any github.com URL, or browse the repos you shared with the Autolab GitHub App. Public repos work with the URL alone; private repos need the App:

  • Configure access installs the App on the account that owns the repository: a personal repo on your account, an organization repo on the organization (an org admin may need to approve).
  • GitHub's install screen is the repo picker. Grant all repos or select a few; change it any time under Adjust permissions or GitHub → Settings → Applications.

Installing the App connects your repos but does not create a climb by itself. After GitHub sends you back, attach the repo and continue.

Models

There is no model to pick. AutoLab runs several agents on your climb at once and rotates among a short list of frontier models, giving more of the work to the models whose attempts improve the score and bringing back a model it has not used lately when the climb gets stuck. Usage is billed to your AutoLab credits.

How a climb searches

A climb runs as a parallel search:

  • Directions. The lead agent proposes a few genuinely different lines of attack (for example "rewrite the solver as a beam search"), marking bold, high-variance ones. An open direction lets work start right away.
  • Workers. Each machine slot hosts a worker: one agent that keeps its own conversation and works one direction. It writes a change, the evaluation scores it, it studies its own result on the machine (logs, profiling) and writes the next change. Every attempt shows up as a scored job.
  • Background searches. A worker can keep a long search (an optimizer, a sweep) running on its machine for up to 12 hours while it keeps submitting experiments, and submit the best results it finds. A worker's running searches share its slot on that machine, and its next steps run there; they stop when its worker retires or the climb stops. This needs a current autolab CLI on the machine.
  • Where machines go. AutoLab keeps an estimate, per direction, of the chance that it produces the next new best, and gives more workers to the directions that keep improving while still trying the others.
  • When it stalls. If many attempts in a row bring no new best, the climb explores wider: workers take riskier ideas, the lead agent rethinks the problem from scratch and proposes new directions, and directions that stopped paying off are retired.

The overview's Search panel shows the directions with their estimates, how many workers are on each, and how each model has done.

In the dashboard, use the climb's sidebar to switch between Overview, Chat, Jobs, Graph, Code, Compute, Memory, Log, and Settings. The header keeps the climb name visible while the next section loads, including if that section returns an error. Opening a job keeps the same climb header; its Jobs breadcrumb takes you back to the job list. Sections and individual jobs have their own links, and browser back/forward works between them.

The overview leads with the current climb state, its explanation, and the available actions. The Auto-generate ideas switch changes that setting without resuming a stopped climb. When enabled, it generates ideas while the climb is running. Model details and compute status are shown below:

  • Stop climb asks how to stop, then you confirm. Let current runs finish (the default) lets running experiments finish and get analyzed, and lets a deep research job that already started write its report; the agent stops coding and nothing new starts. Stop now also cancels running experiments and deep research. Queued work, code, results, and history are always kept, and your own machines stay attached.
  • While runs finish, the climb shows Stopping after current runs finish. Choose Stop now to cancel the rest, or Resume climb to cancel the stop. After Stop now the climb shows Stopping now until the cancelled work has ended on its machine, usually within a minute. When nothing is left in flight, the climb is Stopped. Resume climb works at any point.
  • Idle rented machines go back as soon as you stop. Ones still running an experiment go back when the climb finishes stopping. Stop now releases them at once. Rental billing continues until release.
  • Resume climb picks up where the climb left off. A stopped climb stays stopped after a server restart.
  • Needs input shows the blocker and suggested actions. Enter guidance and choose Reply & resume, choose Resume climb without a reply, or Keep stopped. Keep stopped stops the climb; running experiments finish, but their results wait for you and are analyzed when you resume. The question stays visible as Open question (kept stopped) until you resume.
  • Compute counts current online or starting machines. Rental limit is the maximum the climb may rent, not another machine. Released machines are hidden from this summary. Use Manage for compute settings.

Chat with the AutoLab key is free, including sending and editing messages when your credit balance is zero. Research and rented compute still need credits.

Take it live

init leaves the climb stopped. When the settings look right (autolab settings), take the climb live:

autolab start

The first start seeds main from your code and starts the agent; afterwards it resumes the agent after a stop. Pass --start to init to go live at once. A live climb still runs nothing until it has compute.

Clone and resume

To work on a climb that already exists, yours or one you collaborate on:

autolab projects                 # climbs you can access
autolab clone alice/nanochat     # copy its code into ./nanochat
cd nanochat
autolab status                   # where the climb, and you, are
autolab pull                     # fast-forward to the latest main

You can paste a dashboard link instead of typing the slug, and it also says which control node to clone from:

autolab clone https://app.autolab.ai/projects/alice/nanochat

clone downloads the climb's code, every experiment commit, and links the directory. It does not download logs, metrics, or artifacts; those stay on the control node and are read in the browser. Cloning never changes a climb's state, so a stopped climb stays stopped until you autolab start.

In the projects list, status is the lifecycle (active = taken live) and agent is what the loop is doing now (running, idle, stopping, stopped).

Delete a climb

Owners can delete a climb and everything in it from the CLI:

autolab settings delete           # confirms first
autolab settings delete --yes     # skip the prompt

This is permanent, and it releases any rented compute first. See Climb settings.

Next: Run experiments for the submit/diff/checkout loop, or Compute to attach what makes any of it run.