AchiralAchiral

Concepts · Humans and machines

SOAR uses preferences and a decision procedure to choose one operator from the candidates available in working memory.

Published2026-09-090 reads
Editorial illustration of a brass selector choosing one operator tile from several candidates using preference pins.

Preferences in SOAR

What if several next steps look possible?

SOAR needs a way to choose.

That is where preferences and the decision procedure come in. Productions can propose operators and create preferences about them. The decision procedure reads those preferences and tries to select one operator.

The goal is simple:

text
choose one operator for the current state

Preferences are selection knowledge

A preference does not do the work by itself. It says something about a candidate.

One operator may be acceptable. Another may be better. Another may be prohibited. Another may be required.

The details in SOAR are richer than this chapter needs. The simple idea is enough: preferences help the architecture compare possible next steps.

A small example

Suppose a support agent has three possible operators:

  • retry the password reset
  • ask for the failure code
  • escalate to a human

If the failure code is unknown, the system may prefer asking for it before retrying.

If the account is flagged for security review, the system may prohibit an automatic reset.

If the policy requires human approval, escalation may become required.

Those are different kinds of selection knowledge. They shape the decision before the system acts.

Why this matters

AI agents often hide selection inside a model response.

That makes behavior hard to inspect. If the model chose a tool, why did it choose that one? Was the choice based on policy, history, a prompt phrase, or a guess?

SOAR gives us a cleaner design lesson. Represent the candidates. Represent the reasons for and against them. Then select.

That does not remove judgment. It makes judgment easier to review.

The takeaway

SOAR preferences help choose between operators.

They are not memories of facts. They are not actions. They are selection knowledge.

For AI systems, this matters because tool choice and workflow choice should not be invisible whenever the stakes are real.

Sources

Go back to Operators in SOAR, continue with Impasses and Subgoals in SOAR, or return to the Concepts hub.