An AI agent can still look like a chatbot: inspect what it can do

A chat window does not tell you whether a system only prepares text or can change something in another application. We recommend inspecting the permitted actions and approval points instead of judging by the interface or the word “agent”.

Anthropic's engineering guide distinguishes predefined workflows from agents whose models direct their processes and tool usage. It also describes models augmented with retrieval, tools and memory. Those are implementation choices, not a visual distinction between a chat box and a background service.10

Ask what happens after the answer

Use this recommended action map for a proposed task:

• What approved input can the system read?
• Does it produce a suggestion, create a draft, or write to another system?
• Who checks the source and output before the write?
• Which exact permission allows the action, and can it be narrowed?
• What does the system do if an input is missing, a tool fails or the result of a write is uncertain?
• Which actions can actually be reversed, and who handles those that cannot?

One illustrative difference

Suppose you ask a system to prepare a meeting. A conversational output might suggest an agenda and possible times. A tool-enabled design could also read approved availability. Creating an invitation is a separate action that needs its own authority.

For a conservative pilot, the organiser reviews participants, time zones, clashes and the invitation text before anything is sent. A missing calendar permission or ambiguous date becomes an exception. This is an illustrative design, not an Amulet calendar integration or a tested staffing outcome.

The same separation applies to a document: drafting from an approved source is different from declaring a contract compliant or publishing it. Keep qualified review and external commitments outside the assumed authority of a drafting step.

When not to add an agent

If a template, existing feature or fixed rule handles the task, start there. Anthropic's engineering guidance likewise recommends increasing complexity only when needed.10 Do not justify a purchase with an invented comparison to replacing two staff roles or with a promise of reliable unattended work.

For the fuller permission and acceptance worksheet, see AI agents for knowledge workers. To define one bounded task and its no-build alternatives, discuss the scope with Amulet.

Sources

[10] https://www.anthropic.com/engineering/building-effective-agents — Building Effective AI Agents \ Anthropic
[12] https://support.microsoft.com/en-us/outlook/frequently-asked-questions-about-copilot-in-outlook — Frequently asked questions about Copilot in Outlook | Microsoft Support