Should multiagent systems model their environment explicitly?

Zambonelli, Jennings and Wooldridge's 2003 Gaia methodology argues that the environment a multiagent system operates in — physical ("the temperature in a room, the status of a washing machine... the average speed of cars on a motorway") or virtual ("enterprise information systems, web services, and database management systems") — should be "considered as a primary abstraction during the analysis and design phases," not assumed implicitly. Modeling it means "determining all the entities and resources that the MAS can exploit, control or consume when it is working towards the achievement of the organizational goal."

The paper's argument is that skipping explicit environment modeling, as the earlier 2000 version of Gaia did, "may complicate the overall design and may introduce mismatches between the MAS design and its actual operation." It illustrates this with a conference-management system: without an explicit environmental model, the design would have specific application agents acting "as repositories of papers and review forms," responsible for "explicitly transferring them to other agents" — a solution the authors call "more complex" and "conceptually farther from the likely operating environment," which is actually "a world of Websites from which agents can access papers and review forms." Modeling an explicit environmental data repository instead is both simpler and closer to how the system actually works.

This is an early, methodological version of a claim Can workplace AI risks emerge from interactions alone? makes for a different purpose: that environment, not just an agent's own goals or competence, co-determines what a system actually does. Gaia's reason is design simplicity and fidelity rather than safety, but the diagnosis is the same — a mismatch between the environment as modeled and the environment as it actually operates generates problems that don't show up in any single agent or role. Do single agents always hit organizational limits? extends this to systems with multiple interacting views; Gaia's point is narrower but prior to it chronologically: before any graph of agents can be engineered, the environment that graph sits inside has to be modeled as its own first-class abstraction. Should organizational rules be designed separately from agent roles? is the sibling abstraction from the same methodology: Gaia treats environment, rules, and structure as three separate abstractions that a role model alone obscures.

The excerpt's evidence is a design argument built on two worked examples (a manufacturing pipeline, a conference-management system), not an empirical study — the "complication" and "mismatch" it describes are the authors' own methodological claims, illustrated rather than measured. It does not address environments that are themselves dynamic, adversarial, or language-mediated, which is the situation LLM agents operate in when their "environment" is a tool-calling context window or a live web page; whether explicit environment modeling in Gaia's sense transfers to that setting is outside the scope of this 2003 paper.