Operators in SOAR

What does SOAR mean by an operator?

An operator is a possible next step.

It is not a whole plan. It is not a memory. It is something the system can apply to the current state.

In SOAR, deliberate goal-directed behavior is modeled as selecting and applying operators to states. That is the basic loop.

Propose, choose, apply

SOAR does not begin by choosing from nothing.

First, productions can propose candidate operators. Then preferences help decide between them. Then SOAR selects one operator and applies it.

In plain English:

Here are the possible next steps.
Which one should act now?
Apply it.
Look at the new state.

That loop repeats until the goal is reached or the system gets stuck.

A support example

Suppose an agent is helping Susan with a failed password reset.

The current state says:

goal: resolve reset failure
user: Susan
failure code: unknown
account status: unchecked

Possible operators might be:

• ask for the failure code
• check whether the account is locked
• retry the reset
• send the case to support

The right next step depends on the state. If the failure code is unknown, asking for it may be better than retrying blindly.

That is the value of operators. They make action selection explicit.

Why this matters for agents

Many AI agents act as if every next step should be invented inside the model call.

That works until the task repeats, the policy is strict, or the tool result is messy.

SOAR suggests a cleaner design. Keep the current state visible. Propose possible operators. Choose one. Apply it. Then update the state.

The model may still help. But the architecture does not leave every routine decision floating inside the prompt.

The takeaway

Operators are SOAR's way of making action concrete.

A state says where the system is. An operator says what could change next.

For AI agents, that distinction is useful. A reliable agent should know not only what it remembers, but which next steps are available.

Sources

• The Soar Architecture: Operators and Goals
• Syntax of Soar Programs
• Laird, Newell, and Rosenbloom, SOAR: An Architecture for General Intelligence

Go back to Working Memory in SOAR, continue with Preferences in SOAR, or return to the Concepts hub.