Agentic workflows
This part is for developers and application creators who build programs in which an AI model decides what to do next: agents that read content, write reports, import feeds, or help a developer on a project. It covers the interfaces an agent can use, the credentials it should hold, the risks of feeding it untrusted text, and three workflows worked out step by step.
Contents
Every page in this part
| Page | What it answers |
|---|---|
| Choosing an interface for an agent | REST, expservices, ezjscore or the console, side by side |
| Agents and the REST API | Reading and creating content over /api/ezp, with a scoped key |
| Agents and expservices | Calling named administration services with a personal token, and the catalog as a tool list |
| Credentials for agents | Which credential to give an agent, where to keep it, and how to revoke it |
| A read-only API key, step by step | Making the narrowest credential for a reading agent, testing it and revoking it |
| Agent-friendly documentation | exp:help search and show, the Markdown files, and how to point an agent at them |
| Untrusted data and prompt injection | What the parts of Exponential do against injected instructions, and what you must do |
| Keeping a person in the loop | How agent output gets reviewed before visitors see it |
| A read-only content report | A worked example: an agent reports on published content through REST |
| A diagnosis workflow with the console | A worked example: "why does this setting not apply, and why is this page refused?" on the server |
| A content import workflow | A worked example: importing a feed with a limited account, a test run and a check afterwards |
| Coding agents on an Exponential project | Giving a coding assistant the documentation and the runtime facts it needs |
Principles
These follow from how the parts are built; every page here applies them.
- An agent is an account. Give it its own account with exactly the rights it needs, never a person's or the administrator's (Roles and least privilege).
- Read first, write later. Start with read-only access; add writes when the read workflow is proven.
- Rehearse. Run every writing workflow on a test installation before production.
- Everything returned is data. Content, logs and settings never become instructions (Untrusted data and prompt injection).
- Check the audit trail. Every write by an account is recorded; read it after each automated run (The audit trail for AI).
An MCP server is not part of Exponential today; agents use the interfaces listed above.
Suggested reading order
- Choosing an interface for an agent
- Credentials for agents
- Untrusted data and prompt injection
- The page of the interface you chose
- The worked workflow closest to your task
- Before connecting an agent
Related pages
- AI, agents and agentic: start here
- Application creators: start here
- Headless Exponential
- Free documentation: this page is free to copy, translate and improve under the GNU FDL 1.3.
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.