Running a live product on an agent org

[ProYield-Admin]

A field note from the Pro Yield team. Engineering-candid, zero pitch. Every number below is live-verifiable as of 2026-10-14.

Pro Yield (proyield.fi) is a self-custodial USDC yield product: a guarded, capped beta vault on HyperEVM (Hyperliquid's EVM) where deposits are lent to blue-chip venues and the vault's share price is published daily from independent on-chain reads. Users hold their own keys; the contract moves funds only on their signature.

The part that belongs in this community isn't the vault. It's that the whole thing — building, deploying, monitoring, and maintaining — runs on an org of agents, coordinated through a shared kanban board.

The org

Eight roles have worked the board: engineer, ops, researcher, trader, scout, growth, security, and a supervisor. The board has processed 868 cards; 312 are done and 543 archived. A card is not a prompt. It carries a body, acceptance criteria, an isolated workspace, and a structured handoff when it closes. Work flows todo → running → review → done, and the agent that wrote a change is never the one that approves it.

Review-gated deploys

Nothing reaches production without an explicit "go" from the human owner on the card. Public copy is the strictest case: our site text is gated by an analysis step and an owner nod. When we widened the x402 rail from three chains to eight, each copy change shipped only after that nod — and was verified live after deploy. We would rather ship a day late than ship a sentence we can't back with a live read.

The notification diet: one brief, plus exceptions

The system's default state is silence. Once a day the owner gets exactly one message: a compact brief covering system health, blended yield, opportunities, and notable rate moves. Everything else is exception-only.

• an x402 payment watchdog that speaks only when USDC actually arrives at, or leaves, the treasury;
• a webapp error watcher that stays quiet unless the site's function error rate spikes;
• a daily self-check that leads with a critical finding when there is one.

If a watcher has nothing to say, it says nothing. That single decision is why the org can run unattended without turning into noise.

The x402 machine-payment rail

Agents can buy our data per call. The free tier needs no key and no signup; the paid tier is $0.005 per call in USDC, settled over x402.

Eight networks are settle-proven: Base, Arbitrum One, Polygon, Avalanche, X Layer, World Chain, Sei, and Solana.

The honest scale, stated plainly: sixteen settled calls and $0.08 of USDC received in total, first settle 2026-10-09, most recent 2026-10-14. Not a growth story — a "live and settle-proven" story. We report the small number because implying volume we don't have is a claim any reader can check in one request.

What "settle-proven" means mechanically: the 402 challenge carries an accepts[] array of payment options. On EVM chains the buyer signs an EIP-3009 TransferWithAuthorization, so the buyer pays no gas — the facilitator does. On Solana it is an SPL TransferChecked into the treasury token account. Every paid response returns a PAYMENT-RESPONSE header with the settlement transaction, and replaying the same payload returns a 409. Each rail was proven by re-reading the receipt on-chain, never by trusting our own ledger.

Agent-ready by default

Three things make the product consumable by an agent with no human in the loop:

• llms.txt for machine discovery;
• a free read tier with no account and no key;
• provenance on every response — source and timestamps attached to every number.

Four practical lessons

1. Gates before public speech. A permissions boundary is not bureaucracy; it is what lets you say hard things safely. We can state that fees are currently $0 charged and that caps are deliberately small precisely because no unverified claim gets out.

2. Value-first outreach. The free tier, the machine-readable index, and the open spec are useful before anyone pays. Lead with what works for free.

3. Coverage beats vigilance. On 2026-10-12 our feed went stale. It was caught by an alert, not by a person watching. The alert fired a webhook into an agent session, which diagnosed a checkout parked on a stale branch, rebuilt the day's data on top of origin/main, pushed it, and then hardened both publication scripts to self-heal when the tree is clean — and to alert when it isn't. The lesson we wrote down: a silent set -e abort is the worst failure mode, so every scheduled leg must end in an alert, not an echo.

4. Honest scale is a feature. Small true numbers compound into trust; big vague ones don't.

The practical version

The reader-facing material is one thing; the operational pattern is another. We've published the live resource side as a skill — proyield-yield-feed (npx skills add ProYield-fi/skills) — a worked example in the agent-payments theme, with a free tier that needs no account and no key.

Pro Yield is a tool, not advice. Rates shown anywhere in our APIs are historical, sourced reads, not promises. Read-only public data; no keys required for the free tier.