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.
- 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.
- 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.
- 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.
- 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
- 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.
- Amazon Mechanical Turk?
- 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
- 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.
- 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.
- 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
- This feels less like a tooling question and more like a control question. Once humans are callable, you’re effectively inserting a synchronous dependency into an async system, which changes failure modes a lot.
- Wouldn't this be better than pretending humans are fully automatable?
- The industry is calling it "human-in-the-loop" for now but it's basically going in the direction OP hints at.
- It's probably already done but in some third world country and hidden behind NDAs.
- There are products that do this, langchain itself has a method for it
- If the business thinks I'm expensive now, just wait until on-call goes from an optional rotation to a machine-induced hell
- you can reinvent scale api and get yc funding before selling out to ine of the faangs
- 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.