Tools
When you describe a task in chat, the agent does not guess — it uses tools to work on your project: search the codebase, read and edit files, run tests, or reach a connected app. Each action shows up in the conversation so you can follow along.
Think of tools as the agent's hands: scoped to your workspace, useful for real engineering work, and paused when something needs your OK.
What are tools
Tools are the actions the agent can take on your behalf. You write the goal in plain language; the agent chooses the right steps. Most steps run automatically. Steps that could change your project, run risky commands, or reach the network wait for your approval first.
You do not pick tools manually — you describe the outcome, and the agent handles the rest.
What the agent can do
Those actions map to a few built-in areas:
On cloud agents, the same capabilities apply in a remote environment. What the agent can reach on the network follows your organization's security settings.
Approvals
When the agent wants to do something sensitive, you see a prompt in chat before it runs — for example a file edit outside normal paths, a destructive shell command, a push to a remote, or a call to an external integration.
- Review what the agent proposed — the command, file path, or integration call
- Approve or deny — if you deny, the agent can try a safer approach
- Allow for session (optional) when you trust a repeat action, such as running your test suite
Routine reads and safe edits often proceed without interrupting you. Anything that crosses a trust boundary — deletes, pushes, credentials, or unfamiliar integrations — stops for your decision.
Read terminal commands carefully, especially deletes, remote pushes, and anything touching credentials or global system state.
Extensions
Your team can extend what the agent can do by installing plugins from the marketplace — integrations, skills, and optional automation for how you already work. Administrators set org-wide limits in Model control and Agent security.