Climb settings¶

Settings define what the agent optimizes within the hill's bounds and how an experiment runs. They are the climb-level defaults the agent respects on every experiment.

View¶

autolab settings

Shows the current objective, constraints, run/setup commands, and the assigned LLM key.

Edit¶

Pass any flag to edit those fields, then it prints the updated settings:

autolab settings --description "second attempt with a wider model"

What the project optimizes is its evaluation, shown read-only: a project is pinned to one version of it for its whole life. To optimize a new version, start a new project from the evaluation's page.

Flag	Meaning
`--name`	Display name.
`--description`	Short project description.
`--source`	Source repo URL.
`--coding-timeout`	Optional coding-step time budget in seconds (120–7200). By default there is no budget — the step runs until done. Retried attempts automatically get 2x. Pass `0` to remove a budget you set.
`--watchdog`	Run watchdog `on`/`off` (default on): the agent reviews each running job every few minutes and decides keep-or-kill instead of hard-killing at a fixed timeout.
`--watchdog-interval`	Seconds between run watchdog check-ins (60–3600; default 300). Pass `0` to restore the default.

Clear a field with an empty string

autolab settings --constraints ""

Coding-step time budget¶

Each experiment starts with a coding step: the agent edits code on your node, sanity-checks it, and pushes a branch. That step has a wall-clock budget (default 10 minutes) which includes the time its shell commands take. If your project's environment setup or smoke tests are inherently slow — big datasets, cold caches, busy nodes — raise it instead of letting attempts time out:

autolab settings --coding-timeout 1800   # 30 minutes

Retried coding attempts automatically get 2x the configured budget.

Run watchdog (long jobs aren't hard-killed)¶

While an experiment is running, the agent checks in on it every ~5 minutes instead of killing it at a fixed timeout: it reads the recent logs (each line stamped with when Autolab received it), compares elapsed time against its own estimate, and explicitly decides keep running or kill now with a reason shown on the job page. A healthy training run that outlives its estimate keeps going; a crash-looping, diverged, or no-longer-useful run is stopped within minutes. Silent runs are still killed by the mechanical stall guard, and dead nodes still recover automatically.

autolab settings --watchdog off            # legacy behavior: hard-kill at the estimate
autolab settings --watchdog-interval 600   # check every 10 minutes

Objective, metric & constraints¶

Autolab optimizes a single objective stated in plain language — minimize val_loss, maximize MMLU, or even an LLM-as-judge comparison. The agent reads the metric out of your run's logs and the objective text, so be specific about the number you care about. Constraints are the guardrails: parameter budgets, time limits, "don't touch the tokenizer", and so on.

Stop policy & cost cap¶

By default a project never stops — the agent keeps researching until you intervene, pausing only for the spend cap, an objective that is truly complete or impossible, or a hard blocker that needs you (no usable node, missing credentials, repeated infrastructure failure). To stop earlier, set a termination condition (--stop-policy) and/or a hard spend cap (--max-cost) when you create the project:

autolab init -y --name nanochat --hill alice/nanogpt-10min \
  --stop-policy "stop when val loss reaches 3.0" \
  --max-cost 100

Adjust them later from the dashboard project settings.

AutoLab compute¶

Settings → AutoLab compute is where you choose the machines AutoLab may rent for this climb, and how many of each. Nothing is selected by default.

The agent's recommendation for this evaluation leads, with the rest of the catalog behind show all machines (searchable and filterable). Every price shown is the 1.2x you'd actually be charged, with the provider price alongside it, and the total under the cards is the worst case with everything running at once. Ask the agent again re-reads the evaluation and re-ranks.

Rented nodes are released whenever the climb pauses or stops, and the surplus whenever fewer are needed — see Compute.

Collaborators¶

Managing who can access the project has its own subcommand — see Collaborators:

autolab settings collab          # list members

Delete a project¶

Owners only, and permanent:

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

Changing the objective or constraints steers the live agent on its next experiment. To pause it first, see Drive the agent.

Appearance (web dashboard)¶

The dashboard has dark and light themes and follows your OS preference by default. The small toggle at the bottom of the sidebar cycles System -> Dark -> Light. The choice applies instantly, is saved per browser, and covers every page including sign-in.

Light mode uses white panels on an off-white background; dark mode uses charcoal panels. Titles, descriptions, buttons and form fields follow the same hierarchy throughout the dashboard, including hill and climb pages and setup dialogs.