Cristiano Castelfranchi and Social Power: Why Agents Depend on One Another
Cristiano Castelfranchi grounded artificial sociality in dependence: agents gain social power when others need capabilities, actions, or resources they control.
Castelfranchi argued that sociality begins with dependence
Multi-agent systems can be described as social simply because several autonomous programs communicate, but Cristiano Castelfranchi argued that this is too shallow. In his 1990 chapter “Social Power: A Point Missed in Multi-Agent, DAI and HCI,” he asked what actually makes one autonomous agent socially relevant to another.[1] His answer centered on dependence. An agent has goals but may lack some action, resource, information, or capability needed to achieve them. If another agent controls what is missing, a social relation exists before any message is sent. This reframed coordination from a problem of connecting independent problem solvers into a problem of understanding how autonomy is limited by distributed capabilities and how those limits create power.
Dependence is compatible with autonomy
An autonomous agent can choose its actions while still depending on others for outcomes it cannot produce alone. Social dependence therefore does not negate autonomy; it explains why autonomous actors have reasons to interact.
Power came from another agent’s goals, not from a central hierarchy
Castelfranchi’s idea of social power was relational. Agent B has power over agent A when A depends on B for some goal-relevant condition. Power is therefore not simply an intrinsic property such as processing speed or authority stored inside B. It arises because of the distribution of goals, abilities, and resources across the society.[1] If A acquires another way to achieve the goal, B’s power may disappear. If the goal becomes more important, the dependence may become more significant. This dynamic view suited open multi-agent systems because social relationships could change as agents entered, left, learned new capabilities, or adopted different objectives. Power became something that could be analyzed from the structure of interdependence.
A change in alternatives changes power
Dependence weakens when an agent has substitutes. This makes social power sensitive to redundancy, coalition formation, and the availability of alternative partners rather than fixed by identity alone.
Dependence relations were formalized into recognizable patterns
Castelfranchi, Maria Miceli, and Amedeo Cesta developed the idea further in work on dependence relations among autonomous agents. They distinguished forms such as unilateral and bilateral dependence and examined how more complex patterns could be derived from simpler relations.[2] The formalization connected practical coordination questions to a structural representation. If one agent needs another’s action to achieve a goal, the dependency can be represented even before the agents negotiate. This gives a multi-agent system a social topology: a map of who needs whom, for what, and under which alternatives. Such a map can reveal likely opportunities for cooperation as well as points where one participant can influence another.
Dependence networks made cooperation an architectural property
Later research turned these relations into dependence networks that could support reasoning about cooperation, exchange, and coalition formation. Work on DEPNET showed how a set of agents described by goals, actions, and plans could be analyzed to calculate dependence relations and identify advantageous social interactions.[3] This is different from starting with a fixed team and assuming everyone will cooperate. The structure itself helps explain why cooperation may be rational. Two agents can discover reciprocal dependence, where each controls something the other needs. More complex networks can reveal intermediaries or coalitions. The result is a more realistic account of distributed intelligence in which coordination emerges from patterns of capability and need.
Cooperation can be generated by structure
Agents need not begin altruistic. If each has something the other requires, reciprocal dependence can create incentives for exchange and coordination. Social structure can therefore support cooperation even when goals remain individually owned.
Social action required more than message passing
In later work on social action, Castelfranchi described interference, dependence, coordination, and other social phenomena as relations that emerge among intelligent actors in a shared world.[4] This extended the original power argument. Communication is only one mechanism of social interaction. Two agents can constrain or enable each other through control of resources, shared environments, competing actions, or institutional positions. This matters because many software architectures treat a successful message exchange as evidence that agents can cooperate. Castelfranchi’s framework asks a deeper question: what social condition makes the message meaningful? A request matters because the receiver can do something relevant to the sender’s goals, and because the sender has some reason to expect or induce cooperation.
Trust later became another form of relational capital
Castelfranchi and collaborators also connected dependence networks to trust. If agents must choose partners, they need beliefs about whether another agent is willing and able to provide the required contribution. Later work described being trusted as a form of relational capital because an agent that is selected by others gains opportunities for cooperation and influence.[5] This extends the social-power framework without reducing power to coercion. An agent can become important because others depend on its competence and consider it reliable. In open systems, that status can change with experience. The resulting picture is richer than a static permission graph: social structure includes objective dependencies as well as beliefs about which partners can be trusted to satisfy them.
Objective dependence and subjective trust are different
An agent may objectively need another capability while distrusting the available provider. Conversely, it may trust an agent that is not useful for the current goal. Reliable coordination requires reasoning about both the dependency and beliefs about partners.
The theory supplied a bridge between sociology and multi-agent engineering
Castelfranchi’s work is notable for treating concepts such as power, dependence, trust, and social action as candidates for computational analysis rather than loose metaphors. The 1990 social-power argument explicitly criticized multi-agent research that postulated sociality without explaining its foundations.[1] By grounding interaction in distributed goals and capabilities, the theory supplied a bridge between social-science concepts and engineering representations. This approach influenced work on coalition formation, social reasoning, trust, and normative systems. It also anticipated a problem in today’s AI-agent ecosystems: a system can contain many autonomous components, but meaningful organization depends on knowing which components control resources or capabilities that others require.
Why Castelfranchi belongs in multi-agent history
Cristiano Castelfranchi belongs in multi-agent history because he made dependence a foundational concept for understanding artificial societies. His work showed that social power can emerge from the distribution of goals, actions, resources, and alternatives rather than from a central authority.[2][4] That insight changes how coordination is designed. Instead of asking only how agents exchange messages, designers can analyze why an agent needs another, what leverage that relationship creates, and how alternatives or trust change the structure. Modern tool-using agent systems recreate these conditions whenever one agent controls a database, credential, model, or specialist capability that another needs. Dependence analysis remains a useful way to make those relationships explicit.
Works Cited
- 01Castelfranchi — Social Power: A Point Missed in Multi-Agent, DAI and HCI shop.elsevier.com
- 02
- 03
- 04
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.
Submit a research lead