Documentation
Sage documentation
What Sage is. An AI agent that turns one product URL and one budget into paid, verified testing. It browses the product itself, designs the missions, judges what real people submit, and pays them in USDC on GOAT Network. Three separate model layers do the reasoning; none of them can move money — an on-chain vault computes every amount and enforces every limit.
Where to start
How it works
The five steps end to end, and the two moments a human is actually involved.
Architecture
Three model layers, one settlement core, and why autonomy here is a gate rather than a loop.
The safety model
What still holds if the model is wrong, jailbroken, or offline entirely.
Settlement & proof
How a payout is computed, capped, signed, and published as a receipt anyone can check.
Fund once, then stop deciding
The standing mandate: the agent picks where to spend, never how much, proposes every move first, and waits out your window.
Private payouts
Decided in public, collected in private: the Starknet rail, the claim link, and what the ledger does not say.
Beyond testing — the money lanes
The verify-then-pay machine also runs gigs and milestone grants, composed by hand and compiled with no model in the path; prices can be stated in the Caribbean’s own currencies (J$, TT$, EC$, Bds$ and the rest — converted once at a stamped rate, settled in USDC). Every payout builds the earner’s verified work record, a lender can verify a signed earnings floor without the record publishing a figure, and the Advance facility lends against witnessed inflow — repaid by the waterfall. The system’s outcomes — cost, speed, access, flow — are measured live on /outcomes.
Who it is for
Founders with a live product and not enough real users to know what breaks. You have a URL and a hypothesis about what a first-time visitor should do, but not enough traffic to learn whether they actually get it. Sage turns that into paid missions and pays the testers for you.
It is deliberately narrow. It is not a generic agent platform, a chatbot, or a bug bounty — it turns one product and one budget into verified testing, and it is judged on whether the payouts hold up.
Lenders and programmes. Every payout Sage makes is verified before it moves and anchored to a transaction, so the byproduct is a work record a lender can underwrite without taking anyone’s word for it — published arithmetic over receipts, never a score. Open a real one on the lender view, read the system’s outcomes against the finance track’s own bar on outcomes, and the controls on the money path in compliance.
What it costs and who holds the money
You set the budget and it sits in a vault you own. Rewards are derived from it deterministically, so the sum of every possible payout equals exactly what you funded. Money that is never claimed is recoverable: stopping a campaign revokes the vault and returns the balance on-chain. Sage never holds your keys.
Is any of this real?
Every payout resolves to a public transaction on GOAT Network with a receipt page citing the evidence it paid for. The case study reports the full Stage 2 result — products inspected, submissions judged, how many were refused — with the place each figure can be re-derived from.