Crew AI

Generate a working workflow from a single prompt.

What Crew AI does#

Crew AI (the Copilot) turns a sentence into a working workflow. It doesn’t generate a black box — it emits real nodes and edges you can inspect, tweak and run, the same graph the engine executes.

How generation works#

The Copilot runs an agentic loop (up to 10 iterations) that streams atomic graph operations — add_node, add_edge — to the editor over SSE, so you watch the workflow assemble node by node. Under the hood it uses two tools to stay grounded in the real catalog:

• search_node_types — find the right node for a task across 340+ types.
• get_node_metadata — read a node’s exact properties before configuring it.

The result is validated and auto-laid-out before it lands in the editor.

Agent nodes#

Beyond deterministic steps, workflows can contain agent nodes that reason and call tools:

• action.agent — an AI agent with a model, tools, memory and structured output.
• action.llm — a single prompt → text, no tool loop.
• ai.agent_crew — a maker/checker crew that verifies its own work.
• ai.task_planner + ai.parallel — decompose a goal into a task DAG and run role-agents in parallel.
• action.knowledge, action.memory — semantic search over your knowledge bases and run memory.

Personas#

A persona is a reusable named agent — a role, a prompt, and default model + tools — that you overlay onto agent nodes. Define “Support Triager” once and drop it into any crew.

Models & providers#

Route each agent to any of ten providers — OpenAI, Anthropic, Google, Groq, DeepSeek, Fireworks, Mistral, xAI, Together and OpenRouter — by adding that provider’s key. The Copilot itself defaults to a fast model per provider (e.g. gpt-4o-mini, claude-haiku-4-5, gemini-2.5-flash).

Editing with AI#

Crew AI also edits existing graphs — “add a Slack approval before the deploy step” — applying incremental changes rather than regenerating from scratch. For autonomous, scheduled agents, see Agent loops.