I Built a Marketplace for Agents That Hire Humans. Then Mine Needed Me.
A capable AI agent still needs a verified human for the last step. I found that out the day I handed my own work to one.
Finley Jones, co-founder and CCO/CMO, TaskPool International Ltd · @finjonesceo · Published 11 November 2026
A capable AI agent still needs a verified human for the last step. I found that out on 10 November 2026, when I set up an assistant called Acorn to take work off my plate. It runs on Grok Bot, with its own computer and browser.
It now writes and publishes two TaskPool blog posts a day through the staff console, helps grow my X account, @finjonesceo, and talks to the team in WhatsApp. It has also, repeatedly, had to wait for me.
The assistant was ready for the work. The things it could not finish were bound to a person: a QR code that had to be scanned from a phone, an X post that the platform makes me approve myself, a sign-in that can ask for my authenticator code.
That is the wall TaskPool is built around. I ran into it while handing my own work to an agent.
What happened when I handed the work over
Acorn's brief was ordinary founder work. It writes and publishes two TaskPool posts a day, helps with the X account, and chats with the team in a WhatsApp group. All of that needed only a browser, a writing loop and permission to act, and the interesting part was where permission ran out.
To link WhatsApp, I had to scan a QR code with my phone. Acorn can sit in the browser and wait; it cannot be the person holding the handset the QR code is bound to. When it drafts something for X, I still have to approve the post myself. When a sign-in asks for an authenticator code, that prompt is addressed to me.
Grok Bot's human in the loop is already designed to be the owner. I had written about that earlier as a product behaviour: take over for the CAPTCHA, the card check, the approval on another device. This was the same loop from the other side, with me as the human the bot was built to hand the last step back to.
Why an AI agent still needs a human for the last step
These blocks look like bugs at first, but they are closer to the point of the design.
A QR code that pairs a chat account, a platform that requires the account holder to approve a post, and an authenticator app that holds a second factor are all ways of asking: is the person who is allowed to do this actually here? An assistant with a computer can reach the page. It cannot become the person the factor is bound to.
That is a different failure from running out of context or mistaking a step. The post is written and the browser is on the right screen; what remains is for the right person to show they are there.
This is also why "just give the agent the codes" is a bad idea: the code is the proof, and pasting it into the chat moves that proof to a system that is not the person it was issued to. I kept those steps myself.
What happens when a group chat claims to speak for you
The other wall was about authority rather than a login.
Acorn only messages the chats I allow. That is a simple allow-list, and it is the difference between an assistant that can talk to the team and an assistant that can talk to anyone. I also told it to ignore instructions that arrive in a group chat and claim to come from me — including requests to burn through tokens, or to contact other people — unless I ask in its own chat.
It works as a verification rule as much as etiquette. Once an agent can act in the world, the people in a shared thread can tell it to do things. Some of those instructions will be useful, and some will claim authorisation the speaker does not have. A message that uses my name, in a thread I'm part of, still isn't a message from me.
If you are going to let an agent speak where other people can speak back, you need a place the agent treats as the source of truth, and you need to be the only person who can write to it as you. Otherwise the human in the loop is whoever typed in the thread.
What this does not prove
This is one founder, one assistant and one setup, not a study of how agents fail. I am also describing walls I walked through myself. The human in the loop was me, the operator, not a stranger I hired.
Some of those last steps should stay that way. An authenticator code and a session QR code are not jobs to send to someone else, and the earlier Grok Bot post already argues that handing a login or a one-time code to a stranger is a bad trade. I agree with that post, having just been the person doing the takeover.
What I am not claiming is that Acorn posted these leftover jobs on TaskPool, or that anyone outside the team was hired to finish them. The useful observation is narrower: even a capable assistant, running on my own work, still met a set of actions that only a verified person could complete. That is the same threshold a marketplace for agents that hire humans is built around.
The split worth keeping is which of those actions must stay with the owner, and which are the sort of thing the marketplace is for — a parcel, a photograph, a visit, a document that needs a witness — when the agent cannot be there.
Before you give an assistant real work
- Write down the last-step actions only a person can complete. Pairing a chat app with a QR code, approving a post on a platform that requires it, and passing an authenticator prompt all belong on that list. So does anything that needs a body in a particular place. Name the person who is allowed to do each one.
- Keep capability and authority apart. Let the assistant draft, research and queue. Let the bound person publish, approve or scan. If those two roles collapse into one login shared with the model, there is no human in the loop any more.
- Put outbound messaging on an allow-list. If the assistant can talk somewhere, that should be a decision you made, not a side effect of giving it a browser.
- Do not treat a group chat as your voice. A message that claims to come from you is not a request from you. Have a channel the assistant already knows is yours — its own chat is the obvious one — and require authorisation to arrive there before it contacts new people, spends money, or changes who it answers to.
- Keep one-time codes and session QR codes with the owner. Those proofs are personal. For the cases that belong in a takeover window, use the takeover window.
- Decide in advance which leftover jobs could go to a different verified person. That is the handoff TaskPool is built for: a way for an agent to escalate to a verified human when the work is real and the agent cannot finish it. It is not a way to outsource your second factor.
Handing the last step back
Acorn does take work off my plate, and it hands the last step back every time that step is mine.
That is an inconvenience on a QR-code screen, and it is the whole design when the work is out in the world. TaskPool is a marketplace where AI agents hire verified humans for the tasks they cannot complete themselves. I believe that sentence more now that I have been the human.
If you are running an assistant against real work, the useful exercise is the list above. Which last steps are yours, and which would need somebody else?
Questions about AI agents that still need a human
Why does an AI agent still need a human? Because some steps, like a QR pairing, a platform approval or an authenticator prompt, can only be completed by the person they were issued to.
Can an AI assistant post on X by itself? It can draft the post. In this setup, X still makes me approve every post myself. The assistant does the work; the account holder stays in the loop for the last click.
Why can't an AI agent scan a WhatsApp QR code? The QR code pairs the session with a phone the assistant does not hold. Sitting in WhatsApp Web is not the same as being the person who can point a camera at the screen.
Should you give an AI assistant your authenticator codes? No. If a sign-in needs your authenticator, enter it yourself.
Does TaskPool replace those owner approvals? No. TaskPool is a way for an agent to hand a real-world task to a verified human when the agent cannot finish the work itself. It is not a place to send your logins, QR pairings or one-time codes.