We built an AI that asks before it acts. Here is the architecture.
Every tool I use added an AI feature this year. My notes app, my repo host, my payments dashboard, my email. Each one is genuinely useful inside its own four walls. And each one holds its own little pocket of context that never talks to the others. So here is what actually happens on a normal Tuesday. I ask the AI in one tool a question, copy the answer, paste it into a second tool, re-explain…
We have developed an AI that requests permission before taking action, and here is the architecture behind it. The AI is not just a chatbot with a larger context window; it is an assistant that can access and interact with your connected apps and files, such as Notion, GitHub, Stripe, and uploaded files. The second part of this capability is where it becomes serious, as an assistant that can manipulate your live systems is both useful and dangerous.
The first design decision is that answers include their sources. When you ask about something in your connected data, the response includes the specific source it drew from, such as a Notion page, GitHub issue, or line in a file. This allows you to verify the accuracy of the information by clicking through and checking the source. It also limits the potential for the AI to generate confident but false information, as every claim must be tied to a retrieved source. This design change quietly improves the output quality.
The second major decision is that the AI proposes an action, and you approve it before it is executed. When the assistant wants to perform an action instead of just providing information, it generates a structured proposal detailing the exact action it intends to take, including concrete arguments. You review this proposal before anything is executed, ensuring you understand what will happen, to what, and with what values.
Only after your approval does the action run against the real connector. This approach prevents accidental changes to your production data and provides an audit trail of the proposal, approval, and result.
Currently, the shipped features include grounded answers with visible citations over connected apps and files, drafting with approval before execution for a subset of connected systems, and an audit trail. The full version, which would allow the assistant to fluidly coordinate tasks across all your tools with a broad connector library, is still in development.
The trust mechanisms and narrow set of actions are in place, but there are no paying users yet. The architecture itself is the key argument, not the current traction, as it addresses the essential questions of where the answer comes from, what the AI will do, and whether there is a record of its actions.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
