Escalate: a human on call for when your agent hesitates

Most agents know how to ask for permission, but not how to ask for an opinion. When an autonomous agent reaches a decision about taste or judgment (which headline reads better, whether a tone is appropriate, whether an answer crosses a line), it has no one to ask and ends up deciding on its own. A small experiment presented this week on Hacker News proposes exactly that: letting the agent ask a real person for a second opinion.

The project is called Escalate and describes itself plainly as a human as a service for your agent. If the agent gets stuck or receives a question of taste or judgment, it can consult a flesh and blood person. For now that person is the author himself, as he admits in the writeup, and the source code is available on the landing page.

The pattern behind it

Beyond the anecdote, the idea touches a serious pattern: human-in-the-loop. Today's agents already know how to pause and ask for approval before a sensitive action (deleting a file, sending an email, spending money). That is a permission control: yes or no. What Escalate proposes is different. It does not ask whether it may do something, it asks what it should do when there is no obvious right answer. It delegates judgment, not permission.

That distinction matters because many agent failures are not technical, they are failures of judgment. The model runs the task without trouble, but picks the wrong tone, misreads an ambiguous intent or makes a decision a human would have nuanced.

How it fits with current agents

Anyone working with Claude Code will recognize the mechanics. Tools ask for approval before running actions, hooks let you intercept lifecycle steps and MCP servers connect the agent to external systems. Something like Escalate sits one step above: instead of a yes or no to a specific action, it opens a channel for an open ended question to a human when the agent senses it is outside its zone of confidence.

Technically there is no magic. It is an endpoint the agent sends its doubt to and from which it receives a human answer before continuing. The value is not in the implementation, which is simple, but in formalizing the moment when an agent should stop and ask.

What it does not solve yet

It is fair to be clear about what it is: an early experiment, by a single person, presented as a Show HN with minimal traction. The author himself admits the available human is him. And there lies the obvious limit: a human as a service does not scale like an API. Latency depends on someone being awake, quality depends on who answers, and privacy questions come into play, because the doubt the agent sends may contain sensitive data about the end user.

It also does not solve who the right human is. A taste question about design, one about legal nuance and one about brand tone need different people. Turning that into a reliable service is a much bigger problem than standing up the endpoint.

At ElephantPink we work daily with agents that at some point have to stop and consult us, so the intuition feels right to us: there is no standard way for an agent to ask for judgment, not just permission. Escalate is more a statement of intent than a product, but it points at a real gap in agent design that someone will sooner or later have to fill properly.