Tool calling & agentic execution

What “tool calling” means here

Your agents don’t only generate text—they invoke tools when the model decides an action is needed: search knowledge, escalate to humans, resolve a thread, book time, create decks or PDFs, or call MCP-connected systems (Slack, GitHub, Google Drive, Notion, Confluence, etc.).

Execution stack (how it maps to the product)

1. LLM plans — The model chooses a tool and arguments according to your Agent settings and enabled tools.
2. Runtime executes — The server checks organization, conversation, and session scope, applies rate limits, then runs the tool implementation.
3. Built-in tools — Knowledge search, escalate, resolve, book appointment (see Built-in AI tools).
4. Optional workflow tools — Create deck, create PDF, browser research, website builds, run code, computer use (enable per agent).
5. Custom tools — HTTP or script-style tools you define in agent settings; executed with org-scoped context.
6. MCP tools — Exposed through the MCP integration layer; many paths run MCP servers in an isolated sandbox so credentials stay off the client and access is bounded.

Sandboxed execution (where it shows up)

• MCP — Remote tools run in isolated sandboxes where configured; credentials are injected server-side, not pasted into prompts.
• Website crawler — Fetching and extracting pages may run in a sandbox so extraction is consistent and contained.

For a product-level mental model of managed sandboxes (including patterns popularized by providers such as E2B), see the blog: E2B sandboxes for AI agents.

Together, this gives programmatic execution with clear boundaries. You can review conversation history, approval records, and run traces in the dashboard — not arbitrary access to your core data stores.

“Agentic execution engine” (vision vs today)

The roadmap includes deeper orchestration: multi-step plans, branching, and richer automation. Today, production behavior centers on tool calls (single or sequential turns) plus MCP and custom tools. Workflows add visual pipelines (manual, webhook, and schedule)—see Example workflows for patterns.

Best practices

• Enable only the tools each agent needs (principle of least privilege).
• Use MCP for systems of record; use built-ins for conversation state (escalate/resolve).
• Test in Agent chat before rolling out to Channels.

Related

• MCP overview
• Conversation lifecycle
• Custom tools + MCP validation pack
• Coming soon connections for more systems you can attach when connectors ship
• E2B sandboxes for AI agents — architecture and security patterns for isolated tool runs
• Serverless GPU for model serving — inference scaling when you add custom or private models