Autolab

Autolab offers automated AI model optimization. Customers define a goal and the metric that matters to them — such as accuracy, cost, or latency — along with constraints and an evaluation script that prints the metric. Autolab's agents then read the customer's repository, locate the evaluation, and translate ideas into concrete code changes that can be tested as individual experiments. Each experiment is dispatched to the next available GPU in the customer's fleet, and results feed back into the selection of subsequent experiments.

The system schedules compute continuously: it monitors every job and terminates runs that have plateaued or are clearly failing, releasing that GPU to the next queued experiment so more useful experiments complete without additional compute. Code, metrics, and the full history behind every run are retained, including evidence from stopped runs. At the end of a run, Autolab returns the winning change together with its results and history for human review and approval before anything ships.

The product is delivered as a command-line tool installed via a shell script from app.autolab.ai, with commands to initialize a project against an evaluation script and start an agent run. It also integrates with coding agents, offering installations for Claude Code (invoked with a /autolab command to run experiments on a repository) and Codex (queuing Autolab experiments against an eval script).

Business model

Autolab provides software that runs experiments on the customer's own GPUs, with access currently gated behind an early-access request and demo bookings.

Traction

The product is offered on an early-access basis, with prospective users asked to submit an email address to receive access details.