ForjinnForjinn
Core Concepts

Agentflows: Multi-Step Agent Orchestration

Agentflows let you design multi-step AI pipelines where autonomous agents plan, reason, delegate tasks, and collaborate to complete complex goals.

An agentflow is a visual pipeline built around autonomous AI agents rather than a single prompt-response chain. Where a chatflow follows a fixed path from input to output, an agentflow allows agents to decide at runtime which tools to call, which sub-agents to delegate to, and when to ask for human input. This makes agentflows the right choice for tasks that require reasoning across multiple steps, external data retrieval, or coordination between specialised agents.

How agentflows differ from chatflows

ChatflowAgentflow
Execution modelLinear — nodes execute in a fixed orderDynamic — agents choose their own execution path
Decision-makingDetermined by the pipeline structureMade by the agent at runtime
Best forQ&A, RAG, single-turn tasksResearch, automation, multi-step reasoning
Human-in-the-loopVia follow-up promptsNative support for human input nodes

When to use an agentflow

Agentflows work well whenever the work cannot be fully mapped out in advance:

  • Research agents — an agent searches the web, reads documents, and synthesises a report
  • Task automation — an agent plans a sequence of tool calls to complete a structured task, such as updating a CRM or filing a ticket
  • Multi-step reasoning — an agent breaks a complex question into sub-questions, answers each independently, then combines the results
  • Collaborative agents — a supervisor agent routes work to specialist sub-agents (one for data retrieval, one for formatting, one for validation)

Supported frameworks

Forjinn supports multiple agent frameworks, each with its own canvas and node library. You build agentflows using the framework that best fits your use case.

Each framework has its own section in the Agent Frameworks docs with detailed guidance on nodes, configuration, and deployment patterns.

Creating an agentflow

In the left sidebar, select the framework you want to use: Langflow, CrewAI, AutoGen, or Google ADK. Each opens its own canvas with framework-specific nodes.

Every agentflow needs a starting node — typically a Chat Input or Form Input node — that defines how the flow receives work. Drag one onto the canvas.

Search the node panel for the agent type your framework provides (for example, Agent in LangFlow or AssistantAgent in AutoGen). Configure its model, system prompt, and any tool connections.

Connect tool nodes to the agent. Tools can be built-in (web search, code execution, calculator) or custom functions you define in the Tools section of the dashboard.

For multi-agent setups, add additional agent nodes and connect them to a supervisor or router node. The supervisor decides at runtime which agent handles each part of the request.

Use the built-in chat panel to test your agentflow. When it's ready, save and toggle Deploy in the flow settings to publish a live endpoint.

Executions and state

Agentflow executions are tracked the same way as chatflow executions, with full visibility into which nodes were visited, what tools were called, and how the agent reasoned through each step. Open the Executions page from the sidebar to review past runs.

Each execution record includes the agent's intermediate reasoning steps, making it straightforward to understand why the agent made a particular decision or to diagnose unexpected behaviour.

If an agentflow produces inconsistent results, inspect the execution's Agent Reasoning trail. It shows the exact sequence of tool calls and decisions the agent made, which is often enough to identify where to adjust the system prompt or tool configuration.

On this page