Skip to content

Planning

Planning is configured on the Behaviour node. It is separate from the selected reasoner and from Max Iterations.

ModeCurrent behavior
OffThe agent works without a stored execution plan.
AutoThe model decides whether the request needs a plan.
AlwaysThe model is instructed to create a plan before taking another action.

In Auto mode, the model is instructed to create a plan for work that has multiple distinct steps or dependent tool calls. Straightforward requests can continue without one.

Planning instructions are part of the model prompt. The runtime provides and stores the plan action, but it does not independently reject a different action when a model fails to follow Always mode. Use a model that follows structured instructions reliably and inspect events when planning is skipped.

The runtime supports three plan operations:

  • Create: define the ordered steps.
  • Advance: mark progress and move to the next step.
  • Replan: replace the plan when assumptions or requirements change.

Only one step should be in progress at a time. The current plan is added to later reasoning steps so the model can see completed, active, and pending work.

The plan is stored as an internal fact under the Memory node’s configured scope. It is removed after a final response or error. It remains available across a clarification request so the agent can continue after the user answers.

Current limitation: With User memory scope, the active plan can be visible to another session for the same user. Concurrent sessions can also update the same internal plan. Use Session scope for planned workflows that must remain isolated.

Planning does not increase Max Iterations. A long plan can still exceed the configured iteration limit if the agent uses too many reasoning steps.

Set Max Iterations with enough room for:

  • Creating the plan.
  • Calling and receiving results from tools.
  • Advancing or revising the plan.
  • Producing the final response.

Use Auto or Always for workflows such as:

  • Researching multiple sources before writing a result.
  • Gathering data, transforming it, and sending it elsewhere.
  • Completing a sequence where later tools depend on earlier outputs.

Leave planning Off for simple chat, classification, extraction, or single-tool tasks when an explicit plan would add cost without improving reliability.

If the agent creates plans for simple requests, use Auto instead of Always or strengthen the Behaviour instructions around when planning is useful.

If a plan stalls, inspect the session Events view for tool errors, clarification requests, iteration limits, or malformed plan updates.