Skip to content

Agentic flow

Agentic Flow is the global handover graph opened from the main sidebar. It helps you understand how agents cooperate.

What Agentic Flow is for

Use this view when you want to:

  • understand how one agent hands work to another
  • see the structure behind a shared workflow
  • inspect the path a task can follow
  • review or adjust that structure, if you are allowed to edit it

Understand the handover types

The graph uses named connection types between agents.

  • Delegate sends work to another agent in the background.
  • Consult asks another agent for help and returns to the original agent.
  • Transfer moves the conversation to another agent.
  • Mandatory transfer forces the conversation to continue with another agent.

These labels matter because they change how the user experience behaves in chat and in automation.

An agent whose handover agents have no type yet (configured before handover types existed, or through the API without types) shows each of them as All types: it can delegate to, consult or transfer to each one. Types are set per agent, so giving one of them a type, on the agent's configuration page or on this graph, gives every one of them a type: the others become Delegate, and you are shown which ones and asked to confirm first. Saving the agent's handovers without picking a type keeps them All types, and so does adding a handover agent next to them. Once the handover agents have types, removing the only typed one, or deleting that agent, leaves the others Delegate: they never switch back to All types on their own.

An agent can also delegate to itself: faced with a large workload, it can split the work across background copies of itself that run in parallel, even with no handover agent configured. This is on by default. To keep an agent from spawning copies of itself, turn off Parallel copies of itself in the Agent Handover card of its configuration; it can still delegate to its handover agents, and copies already running finish their work. When an administrator has switched the capability off for the whole platform (AGENT_SELF_FANOUT_ENABLED), no agent spawns copies of itself, whatever this option says.

The global Agentic Flow page

Choose Agentic Flow from the left sidebar to open the platform-wide graph.

This page shows:

  • all visible agents in one graph
  • the handover links between them
  • a warning that changes affect global behavior
  • a focus field so you can isolate one or more agents
  • a graph you can pan and zoom

Use the focus field

Large graphs are easier to read when you focus on a few agents.

Use the focus field to:

  • highlight one agent
  • compare a few related agents
  • reduce visual noise in a large environment

You can add and remove focused agents directly from the filter area, or double-click an agent on the graph. A focused agent that is no longer on the graph (a deleted agent, an old link) is ignored: when none of the focused agents is on it, the whole graph is shown.

Once an agent is focused, choose how the graph shows it:

  • Show only (the default) lays out the focused agents alone: each one, every agent it hands work to or receives work from, and the handovers between them. Double-click one of those agents to add it to the focus and bring its own handover agents in, one step at a time. The counts under the graph are those of what is shown. A handover is created between two agents on the graph: to connect an agent that is not shown, add it to the focus or switch to Highlight.
  • Highlight keeps the whole graph and dims every agent outside the focus.

Open an agent from the graph

You can open an agent directly from the global graph.

This is useful when you want to move from a visual connection view to the full agent detail page.

Editing the global graph

If your role allows it, the global graph is not only a viewer. It can also be an editing surface.

Authorized users may be able to:

  • create a new handover between two agents
  • change the type of an existing handover
  • remove a handover

Because these are global relationships, a change here affects all users who rely on that agent behavior.

Next step

The next chapter explains Work History, where you review runs, waiting items, notifications, and execution details.