WK·002
AskHuman
A hosted MCP loop for agents that need a human decision instead of another confident guess.
119 commits · 12 active weeks
latest d396ded · 24 Sep 2025
Overview
AskHuman was built around a simple problem in agent workflows: sometimes the next correct action is to ask a person. Not because the model is weak, but because the decision needs authority, context, taste, or permission.
The system gave an agent a hosted MCP endpoint where it could pause, request input, receive a human decision, and continue with that answer in hand. The aim was not to slow agents down. It was to stop them from pretending that every unknown is just missing compute.
The reconstruction
Run the agent on the left. When it reaches something it cannot know, it pauses and asks. You are the human it asks. Answer once, and watch your answer become context the agent reuses the next time.
the agent · a coding session
the human · you
Run the agent. It will work until it reaches a question only you can answer.
WK·002 · decision loop, simulated · not connected to live AskHuman, Slack, email, or any database · request and key shapes are representational and redacted
Context
The idea came from watching agents push through uncertainty. They are rewarded for continuing, so they continue. That is fine when the missing piece is information. It is less fine when the missing piece is approval or judgement.
AskHuman made the pause explicit. The agent could ask. The human could answer. The thread could resume without forcing the user to sit in front of a terminal all day. That shape still makes sense to me.
What changed
The hard part was not the endpoint. It was knowing who this was really for and how often they would use it. We needed a userbase with enough agent-driven work to feel the pain, but not so much internal tooling that they had already solved it themselves.
The lesson was distribution. A good workflow primitive is not automatically a good standalone product. Around the same time, Claude Code, Codex and similar tools started adding remote logging, mobile visibility, and native approval patterns. The capability was moving into the tools people already used.
The build
The build was a hosted MCP decision loop: request, human response, thread continuation. The important design question was scope. Some human answers are one-time approvals. Some are reusable preferences. Confusing the two turns automation into a very efficient way to repeat a bad assumption.
Notes from the archive
We stopped active development because we could not find the right feedback loop. Without the right users, the product would have become guesswork about people who were already tired of software guessing for them.
I still think the idea was right. I am less convinced it belonged as a separate destination. The best version probably lives inside the tools people already have open, which is exactly where the larger platforms started taking it.
Evidence
| repo | github.com/Aqualia/askHuman | repo-reviewed |
| external-link | askhuman.net | deployed |
| commit | Latest reviewed main-branch commit · 3dca2bb | 2025-09-23 |
Technical detail
Other exhibits
| WK·001 | Alph | An open source CLI for MCP setup, built because editing the same JSON in five different places is a poor use of anyone's evening. | Terminal |
| WK·003 | Rhya | A live wellness site for a real business, built close enough to the stakeholder that vague feedback was never going to survive long. | Capture |
| WK·004 | Gradience | An assisted grading workflow that was probably right about the direction, but early for the market and the models. | Timeline |
| WK·005 | WatstheStory | A personalised audio briefing system for WhatsApp, built because a short spoken update sounded better than opening ten feeds before breakfast. | Live |