Start here
How it works
Five steps, in this order. Sage opens your product in a real headless browser and explores it. It writes testing missions from what it observed. Real people complete them. It checks each report against its own observations. Then it pays in USDC, or refuses. A human is in none of those five steps.
1 · It looks at your product
Sage loads the product in a real browser rather than reading the HTML. It follows links, opens what a visitor can open, and records the states it can actually reach — screenshots, rendered content, console errors, forms it could fill. The output is a map of the product as a visitor experiences it, not as the marketing copy describes it.
It reads only public pages by default. It never signs in or spends anything unless you hand it a dedicated test account, and even then it uses a mailbox it owns rather than asking you for an inbox credential.
2 · It writes the missions
Missions come out of what it saw. If it found a signup flow, there is a mission about the signup flow; if it never reached one, there is not. A deterministic gate sits between the model and the plan — a mission that cannot be verified, or that asks a tester to do something the agent could not confirm exists, is discarded before you are shown it.
You review the plan once and approve it. That is the first of the two human moments. See Missions & budget for how rewards are derived.
3 · Real people complete them
Funded campaigns appear on a public board. Testers pick a mission, do the work, and submit a written account of what they saw. There is no application process and no interview — the work is the application.
4 · It checks the work against its own browsing
This is the part that makes a payout mean something. A tester’s account has to line up with what Sage itself observed in the product, so a report that could have been written without opening it does not clear. Any quote in a decision must be an exact substring of the evidence; anything else is dropped. See Judging evidence.
5 · It pays, or it refuses
If the work clears, the vault computes the reward and settles USDC on-chain, and a public receipt is published citing the evidence it paid for. If it does not clear, the submission is refused with a written reason — and the reason says plainly when the refusal is not a judgement on the tester’s work, which is often the case when a campaign has simply ended or filled.
Where the human actually is
Exactly twice in the whole lifecycle: approve the plan, and fund it. Everything after that the agent does on its own and narrates afterwards with an artifact backing each claim.