The Agent Guide for Connected AI Agents

Every AI agent you connect to your workspace starts with a detailed guide to what it can do, so it works well from its first request, the same way the docs help your team.

The guide is the same whichever way the agent connects: an AI Assistant connection with the permissions you picked (MountDev Connections), or an AI agent you connected by signing in. It is built from the apps' own settings and from these docs, so it always matches the product. It describes the product and the agent's own connection only. It never contains your records.

What the Agent Is Told When It Connects

When an agent connects, it is told:

• Who and what it is connected to: your one company, by name, and the data region your data is kept in. It acts as you, with only the permissions the connection was given, and never more than you could do on the screens. Every change it makes is marked as made by an AI agent in the Activity Log.
• The apps your company has, one line each.
• How to start: check itself first (below), read the guide, then open the workspace home.
• How permissions work: each tool is one permission at one level (Read, Write, Delete, or Spending), spending quotes first, and some changes wait for an owner or admin to approve them.
• What it can never do: anything it has no tool for, another company's data, the database directly, billing, people, role, or permission changes it was not given, approving or skipping an approval for you, sending an email of a kind that is paused, or guessing at your data.
• What to do when it cannot do something: say so plainly and why, never pretend it worked or work around a refusal, and offer to write up what is missing so you can send it to your provider as a feature request, or a problem report through Get Help in Support when something looks broken.

The Guide's Pages and Ready Jobs

The agent can read the guide's pages whenever it needs them:

• Getting Started: who it acts as, how to plan, and how to read before it changes anything.
• Limits and Approvals: the levels, quotes, approvals, what is saved turned off, paused emails, and the hard limits.
• Errors and What to Do Next: every kind of refusal, what it means, and the next step.
• Turning Repeated Asks Into Workflows: when you ask for the same thing again and again, how it builds a workflow that you turn on (Workflows).
• One page for each app your company has: what the app is for, its records and fields, its common tasks step by step with the tools in order, the errors it may meet, and its limits. An app your company has not added has no page.

It also has ready step-by-step jobs, for the apps your company has: Plan Before You Start, Set Up a New Client (CRM), Follow Up on Overdue Invoices (Invoices), Triage the Helpdesk Inbox (Helpdesk), Build a Weekly Report, and Turn a Repeated Ask Into a Workflow (Workflows). Each one reads before it changes anything and asks you before each change.

Every Tool Described in Full

Each tool the agent may use says what it does, when to use it and when not to, every input with what it means and an example, what it returns, its permission and level, and its common errors with the next step for each.

Plain Refusals

When the workspace refuses a call, the agent is told why in a few words (for example: no permission, needs approval, quote expired, limit reached, or not available in this company) and the next step, so it can explain it to you or use the right tool. The REST API's refusals carry the same why and next_step as well.

The Self-Check

Every connection can check itself (action core.agent.whoami, MCP tool core_agent_whoami, permission core.agent.whoami, level Read). It is always on for every connection and is never a choice when you connect one. It answers with:

• your company's name and data region, and your name and role
• which connection this is
• the apps your company has, and the ones not added
• for each app, the tools this connection has and the ones it does not, with what each would do
• its own calls waiting for an owner's or admin's approval
• the kinds of email that are paused for now
• the limits it works within

It reads only the connection's own setup. It never returns your records, another connection's calls, or anything of another company.