Drive the agent¶
A live project runs an autonomous loop: the
agent proposes experiments, writes code, schedules runs, and analyzes results.
You stay in control of that loop from autolab status.
Where am I?¶
Shows the project at a glance — the agent state (running or stopped), attached nodes, experiment counts, and the experiment you're in. It also prints the project's dashboard link; the browser is where you see what the agent is actually doing right now.
Talk to the agent (browser)¶
The dashboard's Chat tab is a conversation with the research agent itself: ask what it tried and why, what a failure means, or tell it to change direction — "stop tuning the LR, try a different tokenizer". It has the project's full history in context.
Chat is browser-only — there's no CLI equivalent. The CLI's steering controls are the dials below.
Start, stop & resume¶
autolab start # start the agent (first start takes the climb live); resume after a stop
autolab stop # stop the climb; running experiments finish and are analyzed first
autolab stop --now # stop the climb and cancel running experiments
autolab resume # resume a live climb's loop (start does this too)
autolab say "focus on the tokenizer next" # send the agent a steering message (the web chat's channel)
The first autolab start takes the climb live: it seeds the baseline from
your code (or lets the agent write it for an empty climb) and starts the
autonomous loop. After that, start resumes a stopped loop. Pressing Resume
in the dashboard on a climb that never went live does the same first start:
with code on main, it queues the baseline first. Use autolab say to
steer the running agent; the message lands in its context for the next step. A
message you send to a stopped climb is kept, and the agent reads it when the
climb resumes.
Project status vs agent status
These are two different things. The project status is a lifecycle:
setting_up, then active (taken live), then stopped or completed. The
agent status is what the loop is doing right now: running, idle,
stopping, stopped, needs attention, or no llm. That's why
autolab projects and autolab status show both, and the dashboard card
reflects the agent, not the lifecycle.
Stopping is the safe way to make changes (edit settings or adjust the queue) without the agent acting mid-change.
stop starts nothing new and lets running experiments finish. Their results are
analyzed (merged or discarded) before the climb stops, unless you're out of
credits; then they wait and are analyzed when you resume. Results that were
already waiting for analysis stay waiting until you resume. While runs finish,
the climb shows Stopping; it becomes Stopped when nothing is left in
flight. Deep research that is already running completes, but none starts,
including research you request yourself.
stop --now cancels running experiments and running deep research. Both kinds
of stop keep queued work, code, results, and history, and your own machines stay
attached. Idle rented machines go back as soon as you stop; ones still running
an experiment go back when the climb finishes stopping, and --now releases
them at once. A stopped climb stays stopped across server restarts until you
start or resume it.
A climb that keeps hitting an internal error asks for your help (needs input, reason Internal error) instead of sitting idle.
autolab pause still works as an alias of autolab stop.
Share what you know (memory)¶
The climb's Memory tab is what the climb remembers. The agent reads it every time it plans, keeps a short summary of what it knows there ("What AutoLab knows"), and weighs memories from people heavily. Tell it anything it should keep in mind, above all what you already tried before this climb:
autolab memory add -f what-i-tried.md # what didn't (or did) work elsewhere
autolab memory # What AutoLab knows, plus every memory
autolab research "what helps small GPTs converge faster?" # a deep research job: cited report, saved to memory
Deep research jobs show up in the climb's Jobs list next to experiments. See Memory, research and references.
Automatic idea generation (autogen)¶
autolab status autogen # show the current agent + ideas state
autolab status autogen on # let the agent propose new experiments on its own
autolab status autogen off # only run what you queue
With autogen off, the agent still codes, runs, and analyzes — but it only works through experiments you submit. Turn it on to let it explore the search space autonomously. This is the core human-in-the-loop dial: tighten it when you want a hand on the wheel, loosen it to let Autolab run.
Queue your own work alongside the agent with submit, and
watch what it does on the dashboard (autolab open).