Ask HN: When do we expose "Humans as Tools" so LLM agents can call us on demand?

48 points by vedmakk 136 days ago | 31 comments

Serious question.

We're building agentic LLM systems that can plan, reason, and call tools via MCP. Today those tools are APIs. But many real-world tasks still require humans.

So… why not expose humans as tools?

Imagine TaskRabbit or Fiverr running MCP servers where an LLM agent can:

- Call a human for judgment, creativity, or physical actions

- Pass structured inputs

- Receive structured outputs back into its loop

At that point, humans become just another dependency in an agent's toolchain. Though slower, more expensive, but occasionally necessary.

Yes, this sounds dystopian. Yes, it treats humans as "servants for AI." Thats kind of the point. It already happens manually... this just formalizes the interface.

Questions I'm genuinely curious about:

- Is this inevitable once agents become default software actors? (As of basically now?)

- What breaks first: economics, safety, human dignity or regulation?

- Would marketplaces ever embrace being "human execution layers" for AI?

Not sure if this is the future or a cursed idea we should actively prevent... but it feels uncomfortably plausible.

taurath 136 days ago |

I applaud topics like this that get to the banality and dehumanization involved with the promises of an AI future. To me, if AI fulfills even some of its promises then it puts us in a rather high stakes situation between the people that make up society and those that govern the productivity machines.

My first instinct is to say that when one loses certain trusts society grants, society historically tends to come hard. A common idea in political discourse today is that no hope for a brighter future means a lot of young people looking to trash the system. So, yknow, treat people with kindness, respect, and dignity, lest the opposite be visited upon you.

Don’t underestimate the anger a stolen future creates.

impendia 135 days ago | |

Indeed, I wonder if these angry young people would try to fuck with these AI agents, and attempt to make them spin in circles for their own amusement.

Sort of like the infamous GameStop short squeeze of 2021:

https://en.wikipedia.org/wiki/GameStop_short_squeeze

krapp 135 days ago | |

"The child who is not embraced by the village will burn it down just to feel its warmth."

crusty 136 days ago |

Do you work for Peter Thiel and are you tasked with validating his wet dream?

This seems like the inevitable outcome of our current trajectory for a significant portion of society. All the blather about AI utopia and a workless UBI system supported by the boundless productivity advances ushered in by AI-everything simply has no historically basis. History's realistic interpretation points more to this outcome.

Coincidentally, I've been conceptualizing a TV sitcom that tangentially touches on this idea of humans as real-time inputs an AI agent calls on, but those humans are collective characters not actually portrayed in scenes.

econ 136 days ago | |

We will never know how many brain chipped humans are already captured in his catacombs. Perhaps the brain in a jar wetware is further developed than we would like to know.

Now that I think about it. I still do think I'm sitting in my chair.... Humm...

bradchris 136 days ago | |

Check out Mrs. Davis (2023)

crusty 136 days ago | | |

Great show

throw310822 136 days ago |

Easy to do and not a bad idea. You don't need to pass structured output and accept structured input: in the end, an LLM uses any readable text. A tool is just a way for an LLM to ask questions of a certain type and wait for the answer. For example, I'm wondering if certain flows could be improved by a "ask_clarification_question" tool that simply displays the question in the chat and returns the answer.

I understand that this is not exactly in the spirit of your question but, well, a tool is just this.

shahbaby 136 days ago |

When LLMs become better than humans at the following:

1. Knowing what you don't know

2. Knowing who is likely to know

3. Asking the question in a way that the other human, with their limited attention and context-window, is able to give a meaningful answer

victorbjorklund 136 days ago |

Amazon Mechanical Turk?

bobbiechen 136 days ago | |

My thought as well - the infra already exists through MTurk, as well as the ethical and societal questions. You can already pay people pennies per task to do an arbitrary thing, chain that into some kind of consensus if you want to make it harder for individuals to fudge the results, offer more to get your tasks picked up faster, etc.

tmaly 136 days ago |

I sort of see it on the flip side. If you read through the MCP spec, there is the potential for the human input. If should be the AI doing all the grunt work it is capable of with the human putting in judgement when needed to complete some task.

zerocool86 135 days ago |

The framing assumes cloud-first AI agents as the default caller. But there's another path: local-first AI where the human remains the orchestrator and the model never phones home.

The "humans as tools" model only works if the AI layer is centralized and owned by platforms. If inference runs on hardware you control, you're not callable - you're the one calling.

Been thinking about this a lot: https://www.localghost.ai/reckoning

vedmakk 134 days ago | |

Hey, interesting project!

cjmcqueen 136 days ago |

Part of my role is designing assessments for online courses and technical certifications. This is exactly what we want to build in our assessment development process. We want the LLM to monitor the training content and create draft questions and exercises that are vetted by humans. It's maybe a classic "human in the middle" design for content development, but the more we can put humans in at the right point and time and use LLMs for the other parts helps us create a more robust and up to date training and assessment system.

bitwize 136 days ago |

https://en.wikipedia.org/wiki/Manna_(novel)

gnz11 136 days ago | |

Player Piano by Kurt Vonnegut also comes to mind. https://en.wikipedia.org/wiki/Player_Piano_%28novel%29

futuraperdita 136 days ago | |

Exactly this. OP, this is basically where this book goes - AI management that directs humans around as automata.

econ 136 days ago | | |

Right, I could easily write a few prompts to replace all but the lowest level of our orgchart.

jquip 135 days ago |

You don't.

How remarkable to think that humans would be happy to dehumanize each other at least in language, before anything else at the promise of 'optimising' for something... whatever it is... it could be the 200 year futuristic axe at this point.

notjulianjaynes 136 days ago |

I know nothing about this other than I thought it was a joke at first, but I think it's the same idea https://github.com/RapidataAI/human-use

econ 136 days ago | |

That talks about getting some kind of free feedback out of the human for free.

Now we have to find the next level and condition the human to pay to respond to questions.

It seems like an idea bad enough to pay $10 to downvote? Or should that be good enough?

handfuloflight 136 days ago |

Wouldn't this be better than pretending humans are fully automatable?

fromaj 136 days ago |

The industry is calling it "human-in-the-loop" for now but it's basically going in the direction OP hints at.

vedmakk 136 days ago | |

Though human-in-the-loop is usually used in scenarios where control is held by said human (e.g. verification or approval).

The difference I'm curious about is agents being the primary caller, and humans becoming an explicit dependency in an autonomous loop rather than a human-in-the-loop system.

severak_cz 136 days ago |

It's probably already done but in some third world country and hidden behind NDAs.

mountainriver 136 days ago |

There are products that do this, langchain itself has a method for it

bravetraveler 136 days ago |

If the business thinks I'm expensive now, just wait until on-call goes from an optional rotation to a machine-induced hell

htrp 136 days ago |

you can reinvent scale api and get yc funding before selling out to ine of the faangs

muzani 136 days ago | |

Honestly wouldn't mind more competition in this sector. This one doesn't seem optional for the rest of us in the future and I don't like the idea of Scale AI being in charge.

alienbaby 136 days ago |

now the computer decides what it needs, and we bid our time lower and lower to accomplish the task .. :/

Maybe I write a bot that answers fivverr requests at the lowest price possible. We can all race to the bottom.