Build collaborative multi-agent flows with CrewAI

CrewAI lets you build teams of AI agents with defined roles, goals, and tasks that collaborate to complete complex multi-step work in Forjinn's canvas.

CrewAI brings a role-based, crew-oriented model to multi-agent workflows. Instead of a single agent handling everything, you define a Crew of specialized agents — each with a role, goal, and backstory — and assign them Tasks to complete. The crew coordinates work either sequentially or through a manager agent that delegates to workers.

Core concepts

Crew

The top-level orchestrator. A Crew wires together agents and tasks, sets the execution process (sequential or hierarchical), and manages shared configuration like memory and caching.

Agent

A single team member with a Role, Goal, and Backstory. Each agent has its own LLM, tools, and optional memory. Agents can delegate to one another when allowed.

Task

A unit of work assigned to an agent. A Task has a Description and an Expected Output. Tasks can depend on each other — one task's output becomes another task's context.

Tool

A capability available to agents and tasks. Tools can be built-in (web search, code execution), custom functions, LangChain-compatible tools, or MCP tools.

Create a CrewAI flow

In the left sidebar, click CrewAI. The dashboard shows your existing CrewAI flows.

Click Add New and give your flow a name.

From the node palette, add a Crew node. This is the entry point for your flow. Configure the process type, memory, planning, and caching options.

Add one or more Agent nodes. For each agent, fill in:

Role — what position this agent fills (e.g., "Senior Research Analyst")
Goal — what the agent is trying to achieve
Backstory — context that shapes the agent's perspective and tone
LLM — the chat model powering the agent
Tools — capabilities the agent can use

Add Task nodes and connect them to their assigned agents. For each task, fill in:

Description — what needs to be done
Expected Output — what a completed task looks like
Context — other task nodes whose output this task depends on

Wire your agents and tasks to the Crew node. The Crew node takes agents and tasks as inputs and manages their execution.

Add a Crew Kickoff node and connect it to the Crew. This node triggers the execution when the flow is invoked.

Click Validate to check the graph, then Execute to run the flow or Deploy to publish an endpoint.

Crew configuration options

When you configure the Crew node, the key settings are:

Setting	What it does
Process Type	Sequential runs tasks in order. Hierarchical uses a manager agent to delegate work to other agents.
Memory	Enables short-term, long-term, and entity memory shared across the crew.
Planning	Before execution, the crew generates a plan. Useful for complex, ambiguous goals.
Cache	Caches tool results so repeated tool calls with the same inputs don't re-execute.
Stream	Returns output tokens progressively instead of waiting for the full response.
Max RPM	Rate-limits the crew's requests to your LLM provider.

Agent configuration options

Each Agent node exposes a Basic tab with frequently used settings and an Advanced tab for fine-tuning:

Basic settings

Role, Goal, Backstory — the agent's identity
LLM — chat model selection
Tools — tools the agent can call
Memory — per-agent conversation memory
Allow Delegation — lets the agent hand tasks off to other crew members
Verbose — logs step-by-step reasoning during execution

Advanced settings

Function Calling LLM — a separate (often cheaper) model for tool calls
Max Iterations — caps how many reasoning steps the agent takes per task
Allow Code Execution — lets the agent write and run code (safe Docker mode or direct)
Reasoning — the agent reflects and plans before acting
Multimodal — enables image input support
Max Execution Time — sets a timeout in seconds

Task configuration options

Setting	What it does
Output Format	Raw Text, JSON, or Pydantic Model for structured validation
Async Execution	Run the task in parallel with other async tasks
Human Input Required	Pause execution and wait for a human to review the output before continuing
Guardrail	Natural language validation criteria — the agent retries if output doesn't satisfy the guardrail
Context	Comma-separated list of other task node IDs whose results feed into this task

Additional node types

Beyond the core Crew, Agent, and Task nodes, CrewAI flows support:

Knowledge Source — attach document stores or vector embeddings as background knowledge for agents
Memory Store — persistent memory backend for the crew
Flow Start / Flow Listen — entry and event-listening nodes for triggering flows
Flow Router / Conditional Task — route execution down different branches based on conditions
For Each / While Loop — iterate over collections or repeat until a condition is met
State Variable / State Update — read and write shared state within the flow
Safety — filter agent inputs and outputs for harmful content
Grounding — anchor responses to factual sources
Artifacts — handle file and data artifacts produced during execution
Trigger — start the flow from an external event or schedule
Deploy — publish the flow as an API endpoint
Evaluate — score outputs against criteria within the flow

CrewAI flows support Python code generation. After building your flow on the canvas, click Generate Python to export the equivalent CrewAI Python script — useful for running the flow outside Forjinn or inspecting what the canvas is doing under the hood.