Skip to content

2. Agency Structure

Agency as Structure

In agent-centric models, agents communicate, negotiate, and collaborate directly with one another, with the structural aspects of coordination existing within each agent’s internal state. The capacity to organize is predetermined by the agents’ encoded behaviors. While this encapsulation allows the agents themselves to embody the agency, it also means that the agency has no independent representation and is only indirectly observable through the agent's interaction patterns.

Because such agencies are distributed and hard-wired into agents, they are difficult to maintain, modify, and scale. They also tend to create high entry barriers, as any new participant or third-party agent must already possess the embedded knowledge needed to coordinate effectively and preserve system integrity.

The alternative is to create an explicit, externalized representation of the agency framework including its structure, goals, and policies which enables open systems. In such designs, agents, including new entrants, can reason about and adapt to the agency if the representation is accessible and suitably described, supporting greater openness, flexibility, and interoperability.

In agency-centric approaches, agencies are typically conceptualized in terms of their structure — the pattern of information flows, scope of interactions, control relationships, management of resilience, and the distribution of problem-solving capabilities among participating agents. Such a structure provides a framework for interaction by defining roles, behavioral expectations, and authority relationships. In cooperative problem solving, it gives each agent a high-level understanding of how the collective addresses tasks and achieves objectives. However, it can also introduce computational or communication overhead and add complexity to the system.

Agency-centric design focuses on allocating resources to a defined mission or purpose and structuring those resources in a way that effectively achieves the intended goals.

An organization brings together multiple entities that need to coordinate to achieve a goal that none of them could achieve alone. This requires rules that define how the parts fit together, known as the organizational structure. The structure is what remains even when individuals join or leave, as it defines the relationships that turn a collection of elements into a single functioning whole.

In a multi-agent organization, the structure is made up of roles, relationships, and authority lines that guide its behavior. Roles describe the capabilities, objectives, rights, and responsibilities of different members. Knowing their role helps participants set expectations and make plans.

A concrete agency is a specific version of this structure, where the roles are assigned to actual participants, which may be software agents, humans, or other systems.

Agency structures establish communication lines, divide tasks, assign decision-making authority, and provide incentives. They define objectives, roles, interactions, and rules independent of any single participant. The agency’s goals may not be shared by any one member but can only be met through their combined actions. Success depends on bringing in the right agents and arranging their roles, responsibilities, and interactions to support the objectives, as well as on the abilities of the agents themselves.

Social Structure and Interaction Structure

  • The social structure of an agency defines the roles and the dependencies between them. It includes a list of role definitions, group definitions, and a role dependency graph. Roles describe the objectives, rights, and requirements for specific positions.

  • Groups are collections of roles that can be referred to collectively. Norms for a group apply to all roles within it. Groups are defined by an identifier, the set of roles they include, and the norms they follow.

  • In an agency, interactions can be broken into meaningful scenes that follow scripts. Each scene script defines the roles involved, the desired outcomes, and the rules that guide the interaction. The outcome is achieved through the combined efforts of the roles, as each fulfills its objectives or sub-objectives.
  • A scene script also describes the desired patterns of interaction between roles. These patterns can be more or less restrictive, giving agents more or less freedom in how they achieve their objectives and follow norms.
  • Interaction goals are landmarks, which are conditions that must be true at certain points. Landmarks can be arranged in a partial order to create a landmark pattern, which specifies the sequence in which they should be reached. Many different actions can meet the same landmark, allowing agents to choose actions that best match their goals and abilities.
  • Transitions describe how scenes are connected, including the order they occur in and any synchronization rules. Multiple scenes can happen at the same time, and an agent can participate in more than one scene at once. Transitions also set the rules for creating new scene instances and limit how many can exist at the same time.

The Communication Structure

  • Communication in an organization involves both the content of what is being shared and the rules for how it is shared. The content is the domain knowledge, or what the communication is about, while the protocol is how the communication takes place.

  • The content is often described using ontologies, which are shared definitions of terms and concepts in a domain. The communication process itself is usually described using domain-specific languages (DSLs), which define how communicative actions are expressed without depending on the exact content.

  • The communication structure includes both the content and the communication language. Domain ontologies define the knowledge being shared, and communication acts define the language, including the types of messages (performatives) and the protocols for exchanging them.

The Normative Structure

  • At the highest level, norms represent the values of a society. They define what is considered valuable or important but do not tell individuals exactly how or when to act in specific situations. The details of when and how to act are defined in the agency's normative structure.

  • In AgentGrid and AgencyGrid, norms are described in a formal way using PolicyGrid across almost all systems and subsystems of the AI ecosystem so they can apply to specific roles or groups and work under certain conditions and timelines.

  • To enforce norms and handle violations, abstract norms are gradually turned into more concrete norms and then into specific rules, possible violations, and sanctions. This step-by-step process connects high-level values to practical, enforceable actions within the organization.