Skip to content

Agents interact with the outside world through tools. When you ask an agent to search the web, write code, or make a phone call, it uses the appropriate tool behind the scenes. You see every tool call and its result inline in the chat.

Research & web

What you can askWhat the agent uses
"Search for recent articles about AI in healthcare"Web search
"Read this article and summarize it" (paste URL)Web fetch or browser
"Go to this website and fill out the contact form"Browser automation
"Take a screenshot of this page"Browser automation

Code & files

What you can askWhat the agent uses
"Run this shell command: ls -la"Shell execution
"Write a Python script to process this CSV"Python execution
"Create a Node.js script to fetch data from this API"Node.js execution
"Read the report I uploaded earlier"File operations
"Generate a PDF report and give me the download link"File operations

Communication

What you can askWhat the agent uses
"Call this number and ask about my order status"Voice calls
"Reply to that text from Mom"Channel messaging
"When the verification code arrives, enter it"Signals
"Let me know when this finishes"Notifications

Built-in capabilities

Some tools are used automatically by agents without you needing to ask:

  • Memory. Agents store and recall facts about you and their own context. See How Agents Remember Things.
  • Delegation. Agents hand off tasks to specialist agents. See Delegation & Tasks.
  • Scheduling. Agents create recurring tasks. See Scheduling Recurring Work.
  • Signals. Agents pause and wait for an inbound message that matches a description. See Signals.
  • Channels. Agents respond to messages on Telegram, SMS, and other connected channels. See Channels.
  • Notifications. Agents push status updates to your notification feed. See Notifications.
  • MCP servers. Agents call tools exposed by any installed MCP server. See MCP Servers.
  • User interaction. Agents can ask you questions or request your help (e.g., solving a CAPTCHA). This happens automatically when needed.
  • App deployment. Agents can build and deploy web applications. See Building & Deploying Apps.
  • Credential requests. Agents can request access to your stored API keys and secrets. See Managing Secrets & API Keys.
  • Policy management. The system agent can read and edit Cedar policies that govern what every agent is allowed to do. See Policies.

How tools are assigned to agents

Tool access is controlled by policies. The agent settings UI shows a per-agent checklist, where each row toggles a small Cedar policy that allows or denies a tool group for that agent. For finer-grained rules (for example, "only this tool when delegated by the system agent" or "this tool group only when the agent also has X"), edit policies directly.

Some tools are always available to every agent (memory, file output, identity, messaging). The rest are governed by the policy engine.

Sandboxing

Code execution tools (shell, Python, Node.js) run in a sandboxed environment with restrictions on filesystem access, network connections, and execution time. This keeps your system safe even when agents run untrusted code.

See Sandbox for details on how sandboxing works and Policies for how the rules are written.