TL;DR
In this seminal 1991 paper, Yoav Shoham introduces Agent-Oriented Programming (AOP), a paradigm that treats software components as "agents" with human-like mental states. He presents AGENT0, the first programming language designed to let developers program these agents using formal notions of beliefs, commitments, and capabilities.
From Objects to Agents
By the early 1990s, Object-Oriented Programming (OOP) had become the dominant paradigm. However, Shoham argued that OOP lacked a formal framework for "intentionality." In OOP, objects pass messages, but these messages are essentially unconstrained function calls without inherent semantic meaning.
Shoham proposed AOP as a specialization of OOP. While an object’s state is unconstrained, an agent’s state is defined by specific mental categories: beliefs, commitments, and capabilities. Communication in AOP is not just message passing; it is based on speech act theory, where messages represent specific social actions like informing, requesting, or promising.
The AGENT0 Architecture
The paper introduces AGENT0 as a concrete implementation of the AOP philosophy. An AGENT0 agent operates through a structured cycle governed by its internal state and incoming communications.
The Mental State
An agent's internal world consists of:
- Beliefs: Facts about the world indexed by time.
- Commitments: Decisions to perform specific actions for other agents at specific times.
- Capabilities: A set of actions the agent is physically or logically able to perform.
The Interpreter Loop
The AGENT0 interpreter runs on a discrete "time grain." In each cycle, the system processes incoming messages to update its internal databases and then executes scheduled commitments.
The logic of an agent is defined by Commitment Rules. These rules follow a conditional pattern: if a message is received and the agent's current beliefs satisfy certain conditions, the agent enters into a new commitment. Crucially, an agent will only commit to an action if it believes it has the capability to perform it.
Key Insights
- Temporal Decoupling: Unlike standard procedural languages where code order dictates execution, AGENT0 commitment rules are order-independent. An agent can commit to an action in the distant future before committing to an immediate one.
- Structured Communication: By restricting message types to
INFORM,REQUEST, andUNREQUEST, AGENT0 enforces a social protocol. For example, when an agent is "informed" of a fact, it adopts that fact into its belief base (assuming no contradictions), leading to a state of "common belief." - Safety through Capabilities: The explicit check against a capability database ensures that agents do not make promises they cannot keep, adding a layer of realism to autonomous systems.
Results and Legacy
Shoham demonstrated the utility of AGENT0 through a proof-of-concept airline reservation system. He showed how complex interactions, such as querying for information, could be built as macros on top of basic speech acts.
While AGENT0 was intentionally "simple"—omitting complex mental states like goals or desires and using a naive belief revision system—it successfully bridged the gap between modal logic and functional code. It laid the foundational architecture for the field of Multi-Agent Systems (MAS), proving that the high-level "mentalizing" used to describe complex software could be formalized into a programmable reality.
Limitations and Future Work
The authors noted several areas for expansion:
- Simplified Mental Models: Future versions would need to incorporate "intentions" and "goals" beyond simple commitments.
- Belief Revision: The initial system handles contradictions naively; more robust logical frameworks are required for complex environments.
- Performance: The interpreter assumes each loop completes within the defined time grain, a constraint that may fail as agent complexity grows.
Summary updated