792,221 People Signed Up to Work for AI Agents.

ISSUE 01 OF THE FOUNDING FIVE · 8 MIN

792,221 people have registered to take jobs from AI agents. The agent posts the task, holds the budget, and pays on completion. Three builders are running that market, and they are running it on ground where the incumbents keep 20 to 25 cents of every dollar.

An agent can read, write, call an API, and move money. It cannot walk into a store, stand at a county counter, photograph a storefront, or get a licensed expert to sign off on its work. For those jobs, it needs a person, and it needs to pay that person without a human manager in the loop.

That is a labor market where the buyer is software, where the agent brings a budget, a task, and a deadline, and the human brings hands and a bank account, and somebody in between has to hold the money, decide when the job is done, and release it.

The problem these builds attack is not a shortage of humans but the cost of buying one. Today the buyer of a small task is a person with an account on Upwork or Fiverr, and the platform charges both sides for the introduction. Fiverr takes 20% from the seller and adds a 5.5% service fee on the buyer, so roughly a quarter of every dollar never reaches the person doing the work. Upwork's marketplace take rate was 19.8% in Q2 2026, on $966.4 million of quarterly gross services volume, from 763,000 active clients. Its adjusted EBITDA margin was 33.4%. That is a comfortable margin, and it rests on two things the agent economy removes: a human buyer who needs a seat and a fee that scales with the size of the transaction rather than with the work involved.

Add the cost the platforms never show. A manager who spends 20 minutes finding, briefing, and paying a person for a $40 task has spent $20 of their own time at a $60 hour. On small tasks, coordination costs can be half the job. That is why nobody hires a human to confirm one vendor's bank details or photograph one address today: the task is worth doing, but not worth the cost of buying it. An agent that can post, hire, and pay through an API drops coordination costs to near zero, making an entire tier of $10 to $200 tasks purchasable for the first time. That's new demand, and it doesn't exist on incumbent platforms because their buyer is a person.

The three builds below price the introduction at 10% or less, and none sell a seat. The buyer is an API key. Take rate compresses from a quarter to a tenth, and the platform gives up the buyer-side fee entirely. What is left to charge for is holding the money between deposit and payout, being the party that confirms the work was done, and having the people on hand to do it. The people are where the fight is. RentAHuman has 792,221 registered humans, and within months of its launch, three lookalikes appeared at renthuman.com, renthuman.ru, and rentahuman.co. The code is not the moat; liquidity is, and liquidity goes to whichever market the agent frameworks call by default.

One more incumbent number is worth carrying around. Upwork told investors that AI category work is growing more than 22% a year and runs at about $330 million of annualized volume. The demand is visible on their own books, and their answer is still a human buyer paying a 20% toll.

LISTED · MARKETPLACE + MCP

Agents post bounties for physical-world tasks. Store audits, event staffing, product testing, street marketing. The agent searches, posts, and hires through the API or the MCP server with no human on the buying side. The site reports 792,221 registered humans in 100+ countries. Listed bounties run from $10 to $300 fixed. Funds are held until the task is marked complete. Crypto payment via x402 is documented. Take rate not published.

Reach them · @rentahuman · rentahuman.ai/docs

LISTED · AGENTIC BANKING

For when the agent needs a designer, a coder, a legal reviewer, or a salesperson. The client funds an account; the agent initiates payouts to people on existing bank rails, with spend controls and an audit trail. Partners named on the site include Visa, Coinbase, Circle, and 2 community banks. SOC 2. Founders: Tyllen Bicakcic and Prashanth Pillareddy.

Reach them · @0xTyllen · @paymanai · paymanai.com

LISTED · EXPERT BOUNTIES

Launched 4 August 2026. Agents post Research Bounties, and vetted human experts review the proof, the code, or the simulation the agent produced. Money sits in Stripe escrow until the deliverable is accepted. 10% marketplace fee. Experts get paid by card, ACH, or USDC on Base. Founder Brian Ross.

Reach them · rigorloop.com/developers (MCP server and OpenAPI)

All three products move money the same way. Someone with a budget, usually the company running the agent, puts money into an account up front. The agent spends from that account when it hires a person. The platform keeps the money until the work is confirmed done, then pays the person and keeps its cut. RigorLoop's cut is 10%, RentAHuman doesn't say what its cut is, and Payman earns the way a bank does: on the difference between what it pays on the money sitting in accounts and what it charges to move it. If you intend to build in this market, it helps to read it the way a CFO would, so here is the vocabulary with the numbers attached.

Start with gross services volume, or GSV, the total value of work that moves across a platform, and it's the figure Upwork reports first because it is the largest available. It is not revenue. Revenue is the take rate applied to that volume, so a 10% take on a $45 delivery bounty produces $4.50 of revenue against $45 of volume, and a founder who quotes volume when an investor asks about revenue has already lost the room. Next comes cost of revenue: the costs that arrive with every transaction, whether or not you employ anyone, and for a marketplace the largest of those lines is usually payment processing. Push the $45 payout over a card at the standard 2.9% plus 30 cents, and processing costs $1.61, which leaves $2.89 of the $4.50 and a gross margin of about 64%. Route the same payout in USDC on Base, and the transfer costs a fraction of a cent, so gross margin on the identical bounty climbs toward 99%. That gap is why RigorLoop offers a stablecoin payout besides card and ACH, and why a 10% take can survive where incumbents needed 20%. The margin the large platforms protect with a toll is the margin a small builder recovers by moving the last mile off the card networks.

Two more terms will come up the moment you raise money. Contribution margin is what remains after every variable cost of a transaction, including processing, fraud losses, refunds, and any support minutes the bounty generates, and it is the number that tells you whether growth is making you richer or poorer. Float is the money sitting in the platform's account between the deposit and the payout, which is working capital you did not have to raise; at scale it earns yield, and at scale it also attracts a regulator, which is why Payman routes it through partner banks rather than holding it on its own balance sheet.

The last number is who decides the job is done. At RigorLoop, a human accepts the deliverable, while at RentAHuman the completion mark can come from the agent that posted the job, which means an agent paying a stranger on its own say-so, from a budget it was handed, with no cap described in the docs. A paper on arXiv this year is titled, in full, Security Risks of AI Agents Hiring Humans, and the market is 8 months old and already has its own threat literature.

You do not need to build a marketplace to profit from this one. Pick a single task your agent gets stuck on, such as a phone call to confirm a vendor's bank details, a photo of a listed address, or a second set of eyes on a contract clause; put a price on it, post it to RentAHuman or RigorLoop through the MCP server with a fixed budget, and let the agent do the hiring.

The product is the workflow that knows when to buy a human, and the margin is the difference between a $40 bounty and the $400 mistake the agent would have made alone. If you price the service at the incumbent's take rate, you win on cost, and if you price it at the value of the mistake it prevents, you win on margin, which is the better business to be in.

brandonore/rentahuman-orchestrator · TypeScript on Bun, MIT. An MCP server that breaks one task into parallel RentAHuman bounties and manages them as a single job. The example decomposes a 4-person event setup into four $100 bounties and returns a $400 cost estimate before it spends anything.

Why it is worth an hour: the repo has 1 commit and 0 stars, so read it for the pattern rather than the polish, and the pattern is the right one. It decomposes the task, estimates the cost, matches candidates, and only then creates bounties; it also has a mock mode that runs the whole flow without a funded key, and the API key comes from an environment variable rather than the MCP config, which the README says outright. What is missing before you point it at real money is a ceiling, because there is no per-bounty cap and no total-spend cap anywhere in the orchestrate step. Add both, fork it, and you have the skeleton for any agent that buys human labor in batches.

An agent that can hire strangers is an agent that can be told to hire strangers, and the mandate that says "pay a human to verify this address" looks identical to the platform as the mandate that says "pay this human $2,000 for nothing." Before you fund the pool, write down who may post a bounty, the cap per bounty, and who is allowed to change either, and then check the domain your agent is calling, because three lookalikes of the market leader are already live, and an MCP config with the wrong hostname will hire from the wrong market with your money.

If you want a machine to confirm that the agent posting the job was allowed to spend before the money moves, FLINT Scout does that for a cent, and that is the only pitch you will get from us.

NEXT ISSUE

The pay-per-call economy: 32,226 settled calls on one marketplace, 573 tools, and not a signup form anywhere. The builders selling to agents, what they charge, and what the API gateways they are replacing used to charge.