The tools your agents can actually call
By Team Two · Last updated September 23, 2026 · 2 min read
A chat window can draft a sentence. It cannot open the ticket, read the thread, or update the deal unless you paste all of that in yourself. Team Two agents can do those steps, because each agent is given tools.
A tool is a named action with a fixed boundary. The agent does not get a shell on your laptop and a vague instruction to be helpful. It gets the actions you attached to that role, and nothing else.
What ships built in
The built-in set covers the work most teams already do in chat and in the browser:
- Look things up: web search, fetching a page, a calculator, the current date and time, and memory the agent saved earlier.
- Work with files in a sandboxed workspace: list, read, write, and edit. Parse Excel, PDF, Word, and PowerPoint.
- Work with code: Git, code search across indexed repos, Bitbucket pull requests, Playwright tests, and a live preview of a change.
- Work in the systems you connect: Slack, Jira, Notion, Figma, Pipedrive, Intercom, Startdeliver, and email.
- Drive a browser when a site has no API.
Scheduled jobs are a tool too. An agent can run on a clock — a Monday pipeline brief, a nightly check — and deliver the result to Slack or email.
The agent only sees its own list
You attach tools per agent. A support agent can read Intercom and your help docs. It does not also get Git push and a deal-create action because those exist somewhere in the product. A finance agent can parse a spreadsheet. It does not get the sales pipeline unless you add it.
The same rule applies when one agent asks another to help. Specialist agents inherit the caller's permissions, and they cannot spawn further agents. Only the lead agent can fan work out. That limit is there so a side task cannot recruit an ever-growing team. More on who is allowed to see what is in Who is in control.
Built-in tools and MCP are different doors
Built-in tools are the ones Team Two implements and maintains: Jira issues, Notion pages, Slack threads, Pipedrive deals, and the rest of that list. MCP is the open protocol for everything else. GitHub, PostgreSQL, Google Workspace, Sentry, and a server you host yourself show up the same way once connected: as tools on the agent.
If the system you care about is already built in, connect that. If it is not, check the MCP catalog or bring your own server. The map of both is in Connect the stack you already run.
Frequently asked questions
Can every agent run shell commands?+
No. Exec and process tools exist, and they stay off an agent until you add them. File tools are limited to a sandboxed workspace.
Does an agent send email by itself?+
It can search a mailbox and draft a reply. A reply to a customer is not sent until a person approves it.
