Tools are what separate an agent from a chatbot. Grant an agent integration actions, web search, knowledge retrieval, and image generation, and it can reason about a request and then act on it — drafting an email, filing a ticket, or pulling a live answer from the web.
You give an agent its tools in step 4 of the builder. The agent decides which to call, in what order, based on the request and its system prompt.
| Tool | What it lets the agent do |
|---|---|
| Integration actions | Call a connected tool — send a Gmail draft, create a Jira issue, add a Google Calendar event, create a task. Actions come from the integrations your organization has connected, including the one-click OAuth providers and any custom integrations. Integrations. |
| Web search | Look up current information on the open web (powered by the Brave Search API) to ground answers in facts the model wasn't trained on. |
| Knowledge retrieval | Search the knowledge folders attached to the agent and cite the passages it used. Covered in detail in Knowledge & output. |
| Image generation | Produce an image from a text description when the task calls for a visual rather than text. |
Integration actions fall into two groups, and Your Office AI treats them differently for safety. Read actions — fetching a calendar, listing issues, reading a message — run freely because they don't change anything. Write actions — sending an email, creating an event, filing a ticket — have real-world side effects, so they pass through a confirmation gate.
When an agent wants to perform a write action in a chat or agent conversation, Your Office AI pauses and asks you to approve it before anything happens. You see exactly what the agent is about to do — the action, the integration, and the data — and you can approve or reject it. This is implemented natively by the LangGraph engine's interrupt mechanism, so the agent genuinely waits for your decision.
In chat and agent conversations there is a human watching, so write actions are gated and wait for an explicit approval. Inside a workflow, there is no one watching mid-run, and the wiring itself is the consent — connecting an agent node to a write action is a deliberate, deterministic choice. So workflow-driven write actions bypass the gate and run automatically. Deterministic wiring equals implicit consent.
The gate is the safe default, but it can get in the way of a trusted, routine action. For each agent you can mark specific write actions as pre-approved, so the agent runs those without prompting while still gating everything else. For example, an agent might be allowed to create draft emails without confirmation, but still require approval before sending one.
A good rule of thumb: pre-approve low-stakes, reversible actions (saving a draft, adding a label) and keep the gate on anything that sends, deletes, charges, or notifies people outside your team.
Tool use respects your organization's governance. Administrators set a monetary spend cap and a monthly call quota per integration, and every numeric cap has a hard positive minimum — a value of 0 means the integration is disabled, never unlimited. See Integrations for how spend and quota governance works.