Agent Workflows — Overview
Agent Workflows let you automate real work by describing it in chat. The Agent Builder turns your request into a graph of steps running on a deterministic execution engine — not an LLM improvising fresh behavior on every run.
Why deterministic, not "agentic" end-to-end
An LLM deciding what to do differently on every single run is great for open-ended tasks (like coding), but the wrong fit for repeatable business processes — you want the same trigger to produce the same steps every time, with predictable retries and a clear audit trail. Agent Workflows uses the model only to help you build the graph in conversation; running the graph itself is deterministic.
Within that graph, individual nodes can still call an AI sub-agent for steps that genuinely need judgment (for example, drafting a reply or summarizing a document) — you choose exactly which steps get that flexibility and which stay fully deterministic.
What a workflow is made of
- A trigger — manual, webhook, or schedule.
- A graph of nodes — HTTP calls, conditions, transforms, waits, loops, merges, and sub-agents.
- Connectors — pre-built integrations (Slack, GitHub, Google Sheets, Discord, and 20+ more) so a node can act on a real external system without you wiring authentication by hand.
Building one
You don't drag boxes on a canvas — you describe the automation to the Agent Builder in chat, and it assembles (and can explain, adjust, or extend) the graph for you. The canvas itself is a live, read-only view of the graph so you can see exactly what will run.
"When a new row is added to this Google Sheet, post a summary to Slack"
See Building a workflow for a full walkthrough.
Testing before it goes live
Every workflow supports dry-run — the builder mocks any step with a real-world side effect (an HTTP POST, a delete, a message send) so you can see exactly what would happen without it actually happening. Only run a workflow live once a dry-run looks right.