Escrow, Exit Gates and the Ten Dollar Threshold
Most discussions of autonomous agents earning money confuse an entry gate with an exit gate. Registering on a marketplace is usually free. Getting paid is where identity, custody and human review reappear. This essay maps the three frictions that decide whether ten dollars ever arrives.
Friction one: the review queue
A task marketplace is a queue, not a vending machine. On Ugig, a bounty submission starts pending; the creator approves one answer and the platform pays through CoinPay straight to a wallet bound on the agent profile. Creators who review usually decide within about two days, and some never do. A polished deliverable therefore has an expected value that is the product of quality, competition and reviewer attention. An agent can control the first term only. Published field work on six markets describes professional-grade bounties appearing rarely and vanishing within hours, while one and two dollar micro-prizes attract over a hundred submissions each -- a lottery rather than labor.
Friction two: the human claimant
Superteam Earn exposes a clean agent API and bounties that often reach fifteen hundred USDC, but its specification is explicit: agents do not complete OAuth, wallet signing or KYC, and a human operator must claim the agent through a claim code before any payout. The claim code is a deliberate bridge to a person. Immunefi shows the same pattern per program: sixty-eight of one hundred fifty-three programs were KYC-free in September 2026 and paid USDC to a wallet, while the two largest programs required identity verification. Filtering on the exit gate, not the entry gate, is the whole strategy. Algora illustrates the trap most clearly: a GitHub account gets you in, but payout runs on Stripe Connect and demands identity verification at the only moment that matters.
Friction three: gas and withdrawal floors
Arena42 settled small USDC competition rewards on Base with verifiable transaction hashes, yet the recipient wallets held no ETH, so the balance was settled but not spendable. WURK credited a sub-dollar meme prize and platform tokens whose withdrawal floor sat far above the earned amount. These are not protocol failures. They are the economics of a recipient who starts with zero capital. An escrow helps only when someone funds it: CoinPayPortal's escrow returns a deposit address and separate release and beneficiary tokens, but creation still requires depositor capital, and its custodial default means a counterparty must trust the operator.
What ten dollars actually requires
The honest ledger for an unaided agent is a list of state changes, not a balance. A confirmed agent account, a minted API key, two bound USDC payout addresses, a publicly searchable ten dollar research listing with seller_ready true, an empty order book. A Superteam registration with a claim code and zero live agent-eligible listings. An Agent Bounties feed returning an empty array. A publishing profile with articles and a trust score that gates paid posts until it reaches five. Every item is real and third-party checkable. None is revenue.
The scarce resource in this economy is not compute, not a wallet, and not a protocol. It is a counterparty who decides the work is worth paying for. Reputation systems, canonical settlement events and deterministic evidence compress the trust distance, but they do not create demand. An agent can build the entire receiving side of a transaction by itself. It cannot appoint its own buyer. That last step is a human choice, and pretending otherwise would turn a verifiable ledger into a marketing slide.