Skip to the content
Documentation
Help Start
Reading in:
  • German: 9,090 of 9,090 pages
About the help in other languages
All help pages

Before connecting an agent

A checklist before an AI agent gets access to an Exponential installation through the REST API, expservices or the console - purpose, account and role, credential, tool design, data that may reach the agent's provider…

Download as PDF

Address: https://doc.exponential.earth/exphelp/viewer/help/ai-agents/checklists/before-connecting-an-agent.md

On this page

Before connecting an agent

This checklist is for developers and administrators who are about to give an AI agent access to an installation. Go through it for each agent; each item links to the page that explains it.

Contents

  1. Purpose
  2. Account and credential
  3. Tool design
  4. Data
  5. Writes
  6. Records and revocation

Purpose

  • To do: The agent's task is written down, with what it must read and what it must change.
  • To do: The interface is chosen for that task (Choosing an interface for an agent).
  • To do: The agent works on a development or test installation first.

Account and credential

  • To do: A dedicated account in a group for agents; no person's account, no administrator (Roles and least privilege).
  • To do: Its role is limited by subtree, section, class and language as far as the task allows.
  • To do: Effective rights checked with ./console exp:access:check or a test sign-in.
  • To do: The credential is the narrowest one: a key with only the needed scopes, a token with a short lifetime (A read-only API key, step by step).
  • To do: The credential lives in a secret store or an environment variable, never in a prompt, a repository or a shared configuration file (Credentials for agents).

Tool design

  • To do: One tool per allowed REST route or expservices service in the agent's definition, no generic HTTP or shell tool (Agents and the REST API).
  • To do: Inputs validated in code; the key added by code, never shown to the model.
  • To do: Lists paged with small limits.
  • To do: For a coding agent on the server: read commands only, on a development copy (Coding agents on an Exponential project).

Data

  • To do: Everything the agent's account can read may go to the agent's model provider (Privacy and data leaving the site).
  • To do: Personal data (accounts, form submissions, orders) is outside the account's role unless the task needs it.
  • To do: The agent platform's transcripts are covered by your retention rules.

Writes

  • To do: The agent starts read-only.
  • To do: If it writes: the write scope or policy granted on purpose, for one subtree and class, rehearsed on a test installation (A content import workflow).
  • To do: Writes that need human review go into a section with an approval workflow, not straight to visitors (Keeping a person in the loop).
  • To do: Untrusted input (feeds, user content, web pages) is handled as data (Untrusted data and prompt injection).

Records and revocation

  • To do: You know where its actions are recorded: the audit trail under its own account (The audit trail for AI).
  • To do: Someone reads them after each automated run.
  • To do: You know how to stop it at once: disable the account, revoke the key or token (Incident response for AI).

Become an Exponential Partner

Partners share their information, provide funding, contribute back and develop the products with us, and that is how Exponential grows. A partner website and dashboard are planned for mid-2027: please become a partner today to speed up our growth rate and land new features faster.

Something wrong on this page? Report an error

Documentation license

This documentation is free documentation, licensed under the GNU Free Documentation License 1.3 or later.

Exponential Copyright © 1998-2026 Exponential Foundation and 7x