Use cases · Big businesses

Review agent changes on machines you control.

Test instructions on your examples, review the changes and approve each upgrade.

Examples of what you could train. Not customer stories, and not measured results.

Examples

Agents you might train.

After training

What you'd get

  • Evidence for each change: what the reviewers preferred, how the hidden test went, and that you approved it.
  • A record: review the training history and changes to the instructions.
  • Data: training runs on your computer. Prompts and examples go to the model providers you connect.
  • Provider choice: ask the team which model providers your early access build supports.

Day one

How you'd start

  1. Pick one repeated job with clear examples.
  2. Build an agent with real examples of good and bad work.
  3. Run a pilot on a controlled machine with a spending limit.
  4. Review the log and the new version before any use. You approve the upgrade.

Guardrails

Guardrails that matter here

  • Connecting outside helpers is coming soon.
  • You choose the examples and reviewers.
  • Training records let you review each change.

Straight talk

Honest limits

  • It is a single-machine desktop app today, not a hosted, multi-user service.
  • Results have not been published.
  • Rounds are slow today.
  • Ask the team about a build for your computer.
  • Ask the team about licensing, support and data handling before using it in your organization.

Opportunities

Where it can go next.

Faster rounds

Faster rounds are in progress.

Pick one job. Make it better, for real.

Mac and Windows, early access. Free during early access.