Agents that talk, tools that act

An agent is a durable assistant your organization owns. You design it once in Studio. People talk to it for as long as the job takes. The thread stays. The tools stay the ones you allowed.

Studio is where you decide who it is

Write the system prompt. Pick a model that can call tools. Attach prompt skills so policy and product facts merge into the instructions at runtime. Enable built-in tools and named HTTP APIs. Connect outbound MCP servers, probe them, and enable only the tools it should see. Personal data (PII) redaction and moderation sit on the way out, so what leaves the agent is what you are willing to send.

Meet it where people already talk

Tune it in the Studio chat dock. Put it on Slack, Discord, Teams, or email. Keep a thread inside the product. Embed web chat on domains you allow. Each conversation is a session. Your backend can start one, send a message, and stream the reply over Server-Sent Events (SSE).

A pipeline you can draw, or a thread you cannot

Use an ActionFlow when the path is a graph: this, then that, then a person on Slack if the amount is large. Use an agent when the work is still a conversation that has to fetch, draft, and follow through. Same organization. Same credits. Different shape.