A permission-aware multi-agent assistant for a multi-tenant SaaS

The problem

Mitoo is a multi-tenant business platform made up of six product apps. Every customer has its own data, and every user has a role with fine-grained permissions. Everyday tasks like scheduling an interview, logging a sales call or applying for leave take several screens and many clicks.

The goal: let people simply type the task ("schedule an interview for Priya tomorrow at 5 pm") and have it done, with exactly the data and permissions they already have in the product.

product apps

6product apps

specialist agents

28specialist agents

actions it can take

~640actions it can take

common recruiter tasks (was ~1.5 min)

~15 scommon recruiter tasks (was ~1.5 min)

How it fits together

1. Chatinside the product
   • The user types a task in plain language, in the app they're already using.
2. Routerwhich part of the product?
3. Permission layeronly what this user may do
4. Specialist agentone per area of the product
5. Product APIacting as the user

01 / 05

Chat

• ›The user types a task in plain language, in the app they're already using.

What made it hard

• Scale. Hundreds of possible actions across six products. No single model call can choose well among all of them at once.
• Trust. In a multi-tenant system with per-role permissions, the assistant must never be able to do something the user couldn't do by hand, however the request is phrased.
• Exactness. Language models paraphrase, guess and sometimes claim success they didn't achieve. Business actions such as deleting a record or changing a status have to be exact.
• Conversation. Real tasks span several turns: "the second one", "change the date", "yes, go ahead". The assistant has to keep track without mixing up users, companies or apps.

Principles I designed around

• Specialists, not one giant prompt. Each area of the product gets its own agent that sees only what's relevant to it.
• Permissions belong to the system, not the prompt. What the model can't see, it can't be talked into using.
• The model reasons; code handles anything that must be exact. Selections, confirmations and anything destructive are deterministic.
• Act as the user. Reusing the product's own API means isolation, validation and audit trails come for free.

Outcomes

• Common recruiter workflows went from ~1.5 minutes of clicking to ~15 seconds of typing.
• One assistant works across six apps, with each user's permissions respected end to end.
• Recruiters can search candidates by describing the person they want instead of filling in filters.