Reading Tools vs. Doing Tools

Not all tools are created equal. Some tools gather information and hand it back to the agent without changing anything. Others take action — they modify, create, delete, or transmit something in the real world. This distinction is one of the most important in all of AI agent design, because reading is always reversible and doing is often not.

Reading Tools: Gather, Never Change

A reading tool — sometimes called a read-only tool — only retrieves information. It does not modify any data, send any message, or trigger any external action. Examples include: fetching the current weather, searching the web, looking up a word in a dictionary, retrieving a product's price, or reading a file from a folder. The defining property is idempotency: running the same reading tool twice in a row leaves the world in exactly the same state as running it once. No harm is done by calling a reading tool multiple times. If the agent calls a search tool ten times with the same query, the world is unchanged — only the agent's own knowledge has grown. This makes reading tools safe to call speculatively — the agent can gather information before it is certain that action is needed.

Doing Tools: Real Actions with Real Consequences

A doing tool — sometimes called a write tool or action tool — changes something in the world. Examples include: sending an email, posting a message to social media, creating a calendar event, deleting a file, charging a credit card, or submitting a form. The key difference is that doing tools have side effects. Running a doing tool twice does not give the same result as running it once — sending an email twice sends two emails. Deleting a file twice on the first call deletes it; the second call finds it gone. Because doing tools change the world, mistakes are harder to undo. A deleted file may be unrecoverable. A sent email cannot be unsent (usually). A financial transaction may take days to reverse, if it can be reversed at all.

Why This Distinction Shapes Agent Design

Well-designed agent systems treat reading tools and doing tools very differently. Reading tools can be called automatically, in the background, whenever the agent needs information. Doing tools often require additional steps: confirmation from a human, a preview of what will happen, or a log entry so the action can be audited later. This is why many AI assistants ask Are you sure? or show a summary before performing consequential actions. The system is enforcing a pause between the agent's decision and the real-world effect — giving a human a chance to intervene. Think of it as the difference between looking at a map and turning the steering wheel. Looking never hurts. Turning commits you to a direction.