Ask HN: Would you give your Microsoft Azure keychain to an AI agent?
I’m Maxime — a product builder and former Head of Product at Qonto (think Brex for Europe, ~$6B valuation). I recently started something new called [Well](https://wellapp.ai/) (https://wellapp.ai/), where we deploy autonomous agents (via remote browsers or Chrome extensions) to collect supplier invoices on behalf of founders. It saves tons of brain cycles for busy operators.
Now, I know I’m EU-based and this might sound like yet another attempt to regulate everything … but bear with me — the core question is:
> Would you give your Microsoft Azure keychain to an AI agent?
Over the years, I’ve built many integrations — some with OAuth2, others via RPA when no official APIs existed. But with this new generation of agents acting autonomously on behalf of users, I’m starting to wonder: how will we manage authentication and define the scope of what an agent is allowed to do?
*Problem 1: Agent Authentication*
My agents act on my behalf — but I’m extremely anti-password proliferation. While it's tempting to just give an agent my password and 2FA codes, that feels fundamentally broken.
Ideally, I want agents to request access to credentials with a specific scope, duration, and purpose — and I want to manage that access centrally. If I change my password or revoke permissions, the agent should lose access instantly.
*Problem 2: Agent Scope & Consent*
Let’s say an agent gets valid SaaS credentials and starts crawling an account. How do I know it's only collecting invoices, and not poking around in sensitive settings or triggering a password reset?
OAuth solved this with scopes and explicit user consent. But agents today don’t seem to have an equivalent. There’s no "collect-invoices-only" checkbox.
My open question: Should this kind of permissioning live inside a password manager? Or is it the responsibility of agent platforms to build a consent-aware vault? Or should we be thinking about something entirely new — like an MCP (Multi-Agent Control Protocol)?
Would love to hear if anyone has seen serious work or proposals in this space — or if you're tackling similar challenges in your vertical.
Thanks!