The same person can quickly find a transfer route in a subway station with clear signs, yet repeatedly take wrong turns in a station with confusing exit numbers and missing signs. The person’s ability has not suddenly declined; what has changed is how the environment provides information. We can improve navigation by training people to remember routes or by redesigning the signs. Whether the signs work, in turn, depends on whether their users can see and understand them.
The capability an actor actually demonstrates emerges through its interaction with the environment. Understanding and designing an intelligent system that acts continuously therefore requires modeling the agent, the environment, and their interactions together.
Modeling them together requires expressing which information in the environment the actor can perceive, which actions it can take, what those actions change, and how the results feed back into the next round of judgment. These relationships need to connect the actor’s capabilities with environmental conditions so that we can identify where a difficulty lies and which part to improve. As the actor gains experience and its actions change the environment, those connections also need to be kept up to date.
When we consider how agents can enter today’s work systems, this perspective reveals a problem that has long been handled by people and is therefore easy to overlook.
Enterprise software, forms, and processes often function because people continually supply interpretation and coordination. Does “procurement complete” mean that an order has been placed, goods have arrived, or goods have been accepted and entered into inventory? Do the materials in two spreadsheets belong to the same batch? Who needs to adjust their plans when a delivery date changes? People familiar with the business may resolve these questions as they arise, without turning the answers into stable definitions and relationships in the system. The fact that people can keep a business running does not mean its systems explicitly represent the logic those operations depend on.
For an agent to judge whether an order can be delivered on time, it needs to know which materials and production tasks the order involves, their current states, and which dependencies could affect delivery. If these relationships have never been explicitly recorded, switching to a stronger model cannot guarantee that it will reconstruct the same set of business facts every time. What deserves improvement here includes object definitions, how states are represented and maintained, and dependency relationships—not just the agent’s reasoning process.
Some matters go further: the business itself has not yet settled them. Different departments may disagree on what “complete” means, and there may be no agreement about who decides on exceptions. These gaps require business decisions by the relevant actors. A model can identify ambiguity and propose options, but an interpretation it supplies on its own cannot simply be treated as consensus already reached by the organization.
If every agent fills these gaps within its own context, their judgments may rest on different premises even when individual tasks can be completed. One agent may consider procurement complete while another is still waiting for materials. More message exchanges can help them discover disagreements. Resolving those disagreements reliably also requires making their shared objects, states, and rules explicit and deciding who maintains them.
Agent design therefore needs to answer questions on the environment side as well: which facts the system will continually provide, which rules the business must define first, which judgments to delegate to agents, which results require human confirmation, and how execution will update the shared state. These choices determine how much of the burden of understanding and judgment an agent actually needs to carry.
The value of modeling agents and environments together is that it lets us judge which difficulties deserve stronger intelligence and which should be removed by reorganizing the environment. When the same structure has to be inferred repeatedly, or the same ambiguity explained again and again, making it explicit and maintaining it is itself an improvement to the entire intelligent system.