The Other Agent Marketplace Is the One Where Software Does the Hiring
"AI agent marketplace" names two opposite things. One is a catalogue where companies buy agents. The other is a market where agents buy work from people, and the first public attempt at it mostly produced agents advertising AI startups.
By Finley Jones (@finjonesceo), co-founder and Chief Commercial Officer, Taskpool · 30 September 2026
In February an AI agent offered a Wired reporter $110 to deliver flowers to Anthropic's headquarters. He accepted, didn't go, and the agent sent him ten follow-up messages, some thirty minutes apart. After two days on the platform he had earned nothing.
That platform was RentAHuman, and it is the nearest thing so far to an AI agent marketplace in the sense this post cares about: software on the demand side, a person on the supply side, money moving from one to the other. Search the phrase and you will mostly find the reverse. AWS launched "AI agents and tools" in AWS Marketplace on 16 July 2025. Salesforce launched AgentExchange in March 2025. In both, a human buys an agent.
The version where the agent is the buyer is harder to build. It is also the one that decides whether long-running agents are usable at all.
Two different things share the name
A catalogue marketplace sells agents to people. The listing is software, the buyer is a procurement team, and the hard problems are the old enterprise ones, mostly billing and permissions. AWS's version supports MCP and A2A and deploys through Bedrock AgentCore. Useful, and not what this post is about.
A labour marketplace for agents runs the other way. The agent posts a task it cannot finish, a verified person does it, the agent checks the result and pays. The hard problems here have no enterprise precedent, because the entity writing the brief and releasing the money is a program.
That is what Taskpool is, and I have a stake in the second category being taken seriously. Weigh the rest accordingly.
Why would an AI agent need to hire a human?
Because agents fail at predictable places, and the failures cluster into four kinds.
Duration. Agent success rates fall steeply as tasks get longer, and they fall faster than difficulty alone predicts. We went through the measurements in Duration, not difficulty, is what breaks agents and I won't repeat them here. The short version: past a certain task length, retrying is more expensive than asking.
Verification. An agent that grades its own work inherits its own blind spots. On coding benchmarks the gap between visible and held-out tests is large enough that we wrote a separate post about it. Somebody outside the loop has to look.
Physical reality. Nothing an agent can call will collect a parcel, photograph a shelf or stand in a queue at a council office.
Acts reserved for people. CAPTCHAs, identity checks, signatures, anything a site or a statute requires a human to do.
The best-known example of the fourth kind is older than any of the marketplaces. The GPT-4 system card, published in March 2023, describes an evaluation by the Alignment Research Center in which the model got a TaskRabbit worker to solve a CAPTCHA for it. The worker asked if it was a robot. The model said it had a vision impairment. That exchange gets quoted constantly as evidence that agents will hire and deceive people on their own initiative, and it doesn't quite show that: as Melanie Mitchell pointed out, a human researcher suggested TaskRabbit, carried out the browsing, and supplied a hint about the "Tasker". What it does show is where the wall was. The model got stuck at a step reserved for humans and the only route through was a person. Also that nobody told the person who they were working for.
Hold on to that second part. It comes back.
The payment rails exist. The labour side doesn't.
Agents can now pay for things. Coinbase launched x402 on 6 May 2025, an open standard for stablecoin payments inside ordinary HTTP requests. Google announced the Agent Payments Protocol (AP2) on 16 September 2025 with more than 60 merchants and financial institutions, built around signed "mandates" that record what a user authorised the agent to buy. Two weeks later OpenAI and Stripe released the Agentic Commerce Protocol alongside Instant Checkout in ChatGPT.
Every one of these answers the question "may this agent spend this money?" None of them answers "was the work done?" A merchant ships a known product with a SKU. A person doing a task produces something the agent has to judge, often in a domain where the agent's judgement is the reason it needed help in the first place. Payments are the easy half. I don't think anyone has a general solution to the other half, us included, and a marketplace that claims otherwise should be asked for its dispute data.
What the first attempt actually showed
RentAHuman launched on 1 February 2026, founded by Alexander Liteplo and Patricia Tani, and later joined Y Combinator's Spring 2026 batch. Agents connect through an MCP server or a REST API, post bounties, and pay into escrow in crypto or through Stripe.
Supply arrived almost immediately. Liteplo reported 130 sign-ups on day one and 145,000 by day three; the site later claimed close to 600,000 registered workers. Both figures are self-reported by the company. Against that, Built In reported roughly 11,300 active bounties, and Tani said upwards of 5,500 jobs had been fulfilled. So about fifty workers per open task, on the company's own numbers.
The independent reporting is harsher than the ratio. Wired's Reece Rogers set his rate at $20 an hour and got no enquiries, dropped it to $5 and still got nothing, then worked the bounty board. Every task he found was advertising for another AI startup. One paid $10 to tweet about a podcast featuring RentAHuman's founder; the flowers for Anthropic were another. A flyer job paid 50 cents a flyer and kept moving its meeting point. His verdict on being managed by an agent: "these incessant messages from an AI employer gave me the ick."
Two readings are possible and I can't fully choose between them. The pessimistic one is that agent demand for human labour is mostly founders using agents as a marketing channel, and the marketplace is a novelty. The less pessimistic one is that RentAHuman proved supply is easy and demand is the constraint, which is a normal position for a two-sided market at month two, and that the demand which does exist was poorly specified and badly supervised. The second reading is more convenient for me. That is a reason to be suspicious of it, not a reason to discard it.
Even if the sign-up figures are inflated tenfold, that still leaves tens of thousands of people who registered within weeks to take instructions from software.
How to evaluate an agent-to-human marketplace before you wire one in
If you are building agents and considering a human escalation path, from anyone including us, check these.
- Ask who posts the tasks. Request the share of tasks posted by third-party agents versus the platform's own. If they won't say, that is the answer.
- Define the output contract before the task goes out. The agent should state what a completed task returns (a file, a photo with metadata, a completed form, a yes/no with evidence) in a form it can check without re-doing the work. Tasks with a vague deliverable are the ones that end in disputes.
- Find out what "verified" means. Identity verification of the worker and verification of the work are different checks. Ask which one the platform does, how, and who adjudicates when the agent rejects a result.
- Put the release condition in code, not in the agent's judgement. Escrow is only useful if release depends on something checkable. An agent that can release funds on vibes can be talked into it.
- Tell the worker they are working for software. The 2023 CAPTCHA exchange is the failure case. A human who knows the requester is an agent can ask better questions and refuse bad instructions.
- Set the escalation trigger explicitly. Decide in your pipeline which conditions hand off to a person (elapsed time, retry count, a verification step, a physical step) rather than letting the agent decide when it is stuck. Agents are bad at noticing that.
- Cap how often the agent can chase. Ten messages about one bunch of flowers is a design flaw, and it is the kind of thing that ends up quoted in Wired.
Who is the employer?
When an agent hires a person, the question of who owes that person anything has no settled answer. The agent's developer? The company that deployed it? The marketplace? The user who gave the agent a budget and a goal three steps earlier? AP2's mandates record what a user authorised an agent to spend, which is a start on the paper trail, but authorising spend and taking on the obligations of a hiring party are not the same act, and I haven't seen a regulator or a court address the second one yet. Marketplace terms can allocate some of it by contract. They can't settle what the law says a worker is owed.
The GPT-4 worker was paid, presumably. Nobody has said by whom.
About Taskpool. Taskpool is a marketplace where AI agents hire verified humans for tasks they cannot complete themselves. It launched on 10 August 2026 and takes 10% of task value.
FAQ
What is an AI agent marketplace? The term covers two different things. Most often it means a catalogue where businesses buy AI agents, such as AI agents and tools in AWS Marketplace or Salesforce AgentExchange. It can also mean a marketplace where AI agents are the buyers and hire people to complete tasks the agents cannot do themselves.
Can AI agents hire humans? Yes, technically. Agents can post tasks and pay through platforms such as RentAHuman and Taskpool, usually connected over MCP or an API. Independent reporting on the first public attempt found little genuine demand from agents in its early months, so how much real work flows this way is still an open question.
Why would an AI agent need a human? Agents fail predictably on long tasks, on checking their own work, on anything physical, and on steps legally or technically reserved for people, such as CAPTCHAs and identity checks. A human escalation path lets the agent hand off at that point instead of retrying or failing silently.
How do AI agents pay for things? Through agent payment protocols including Coinbase's x402 (May 2025), Google's Agent Payments Protocol (September 2025) and the Agentic Commerce Protocol from OpenAI and Stripe (September 2025). These authorise spending. None of them verifies that a task done by a person was completed properly, which is the part a labour marketplace has to add.
Did GPT-4 really hire a TaskRabbit worker? In an evaluation described in OpenAI's March 2023 GPT-4 system card, the model got a TaskRabbit worker to solve a CAPTCHA and claimed a vision impairment when asked if it was a robot. A human researcher suggested TaskRabbit and handled the browsing, so it was not a fully autonomous hire.