Is PRS the same as a fixed workflow?

No. It can select relevant procedures and reconsider activity as beliefs and goals change, although it still depends on represented procedural knowledge.

Is PRS a belief - desire - intention architecture?

It is an important architecture associated with that approach, using beliefs, goals or desires, and intentions explicitly.

Does PRS require a language model?

No. Its foundational work predates modern large language models and uses symbolic procedural representations.

AI Glossary / Data engineering and architecture / Advanced

Procedural reasoning system

Definition

A procedural reasoning system, or PRS, is an AI architecture for selecting and carrying out procedures while responding to a changing environment. It combines represented beliefs, goals and current intentions with procedural knowledge, allowing reasoning and action to be interleaved rather than requiring a complete fixed plan before execution.

What it means

The term refers to a particular tradition of AI architecture, developed through work including that of Michael Georgeff, Amy Lansky and François Félix Ingrand. It is not simply another name for any program that follows procedures.

A PRS represents what the system currently believes, what it is trying to achieve and which courses of action it has adopted. Its procedures describe ways to pursue goals or respond to situations. These representations support reconsideration when new information arrives.

Belief and intention are technical descriptions of stored information and execution commitments. They do not imply consciousness or human-like experience. A belief can also be wrong or outdated if the system's observations are incomplete.

How it works

Procedural knowledge is organized into units often called knowledge areas, or KAs. A unit specifies when it is relevant and a body describing how to proceed, potentially through tests, actions and subgoals.

The interpreter considers current beliefs and goals, identifies applicable procedures and selects work to pursue. Executing a procedure can introduce subgoals, which in turn require further procedure selection. Observations can update the belief store and make new procedures relevant.

This produces an interplay of ongoing commitments and reaction. A system need not abandon every intention whenever anything changes, but it can reconsider a course of action when a relevant event or failure occurs. Selection and priority policies are therefore important parts of the application.

The architecture does not by itself guarantee that procedures are correct, that observations are timely or that every deadline is met. Those properties require suitable domain knowledge, implementation and analysis.

Three practical examples

A robot-control study represents a delivery goal, available navigation procedures and responses to a newly blocked route.
A supervisory system selects a known recovery procedure when an observed equipment condition makes that procedure relevant.
A research workflow executive maintains an analysis goal while choosing an alternative procedure after a required service becomes unavailable.

Think of it this way

Think of a coordinator using a library of operating procedures while monitoring the actual situation. The coordinator keeps the main objective in view, carries out the next appropriate step and revisits the plan when its assumptions change.

The analogy captures procedural knowledge and reconsideration. PRS does not acquire a human coordinator's judgment merely by naming its records beliefs and intentions. Its behavior depends on the procedures, observations and decision mechanisms that have actually been represented.

Worked example

Selecting a route procedure after a change

In an invented robot example, the goal is to deliver a sample to room R. The belief store initially says corridor A is open. The system selects a procedure that uses A and adopts the next navigation subgoal.

Before entering A, a new observation records that the corridor is blocked. Continuing the old procedure without checking its applicability would conflict with the updated belief. A represented response can suspend that route intention and consider an alternative through corridor B, if B is believed available.

If neither route is available, the system needs a defined alternative, such as waiting or reporting that it cannot proceed. The label PRS does not invent that response automatically. This trace illustrates interaction between goals, beliefs and adopted procedures; it is not executable robot-control guidance or a claim about guaranteed safe navigation.

Illustrative case study

A fictional laboratory automation team builds an executive around reusable procedures. Its first design retries an unavailable data-transfer step indefinitely, keeping the overall experiment goal active but making no useful progress.

The team adds an explicit failure condition and a procedure that records incomplete transfer, preserves the available result and requests the appropriate follow-up. It also distinguishes the belief that a transfer was requested from evidence that the destination received the data.

The lesson is that adaptive execution requires well-defined observations and failure handling. A library of successful-path procedures is not enough to support dependable behavior in a changing environment.

Common misconceptions and mistakes

PRS must construct every action sequence from first principles. It can select and combine represented procedural knowledge.
Reactivity means having no continuing plans. Current intentions give continuity while relevant changes can trigger reconsideration.
Naming the architecture establishes a real-time guarantee. Timing and correctness depend on the implementation, workload and represented control policies.

Business use

PRS is relevant to the history and design of autonomous systems, especially where goal-directed activity must coexist with reactions to events. It offers useful questions for modern agent designs even when they do not implement PRS directly.

Ask what the system believes, how those beliefs are updated, which goal is active and why a particular procedure was selected. Make interrupted, failed and completed intentions distinguishable. These distinctions help explain behavior and expose missing recovery paths.

For an AI-enabled workflow, a language model might help propose procedural content, but generated steps still need validation and an execution mechanism. A prompt that says “follow the goal and adapt” is not equivalent to a PRS architecture with explicit procedural representations, applicability conditions and intention management.

Check your understanding

Why should the robot not treat “corridor A was open when planning began” as enough reason to continue after receiving a blocked-corridor observation?

Answer: Procedure selection and continued execution depend on relevant current beliefs. The new observation changes the situation and calls for the represented reconsideration or failure response.

Related terms

Reactive planning
Execution monitoring
Autonomous agent
Belief–desire–intention architecture
procedural knowledge
knowledge area
goal management
intention

Learn next

Read reactive planning and execution monitoring. Describe a small goal with two alternative procedures, state their applicability conditions and trace what happens when a required condition becomes false during execution.

Sources

Georgeff and Lansky, Reactive Reasoning and Planning: https://cdn.aaai.org/AAAI/1987/AAAI87-121.pdf. Primary paper; relevant architecture, knowledge-area and interpreter passages reviewed.

Ingrand, Chatila, Alami and Robert, PRS: A High Level Supervision and Control Language for Autonomous Mobile Robots: https://homepages.laas.fr/felix/publis-pdf/icra96.pdf. Primary research on procedural supervision.

Ingrand and Despouys, Extending Procedural Reasoning toward Robot Actions Planning: https://homepages.laas.fr/felix/publis-pdf/icra2001.pdf. Primary research on planning extensions. The route trace and laboratory case are original illustrations.

Previous: Procedural memory Next: Processing Bias