Multi-Agent Confidence Propagation

Mechanism

In a multi-agent system comprising a plurality of semantic agents operating in a coordinated executive graph, the confidence values of individual agents propagate through the executive graph and influence the confidence computations of other agents. Confidence is a first-class computed state variable: each agent occupies a designated confidence field that encodes its assessed sufficiency to continue executing its current task, and that the confidence governor uses as a hard gate on execution. Multi-agent confidence propagation extends this single-agent gating so that the confidence state of the system as a whole is informed by the confidence states of its constituent agents, and so that the failure or suspension of one agent's execution is visible to and accounted for by related agents.

The propagation is not a free-floating signal that any agent may rewrite at will. It is structured by the relationships in the executive graph: delegation relationships between parent and child agents, and peer relationships between collaborating agents. Each relationship type carries confidence in a defined direction with a defined effect on the receiving agent's own confidence computation.

Upward Propagation From Child to Parent

When a parent agent delegates a sub-task to a child agent, the parent agent's confidence computation incorporates the child agent's reported confidence as an input. If the child agent's confidence drops, indicating that the child is encountering adverse conditions in executing the delegated sub-task, the parent agent's confidence is correspondingly reduced. The child's confidence value is thus not merely the child's private state; it becomes an input to the parent's own assessment of whether the parent should continue to act.

The magnitude of the parent's confidence reduction depends on the criticality of the delegated sub-task to the parent's overall task. A confidence drop in a child agent performing a critical sub-task produces a larger confidence reduction in the parent than a confidence drop in a child agent performing a peripheral sub-task. This criticality weighting ensures that the parent's confidence reflects the structural importance of each delegated piece rather than treating all delegations as equally consequential.

Downward Propagation of Suspension

Confidence propagation is directional. Child agent confidence propagates upward to parent agents, and parent agent execution suspension propagates downward to child agents. When a parent agent's execution is suspended, all child agents executing delegated sub-tasks on behalf of that parent receive a suspension signal and enter their own suspension procedures.

The child agents do not suspend uniformly. Each child enters its suspension procedure according to the task class differentiation that governs confidence interruption: a child executing a terminal sub-task preserves state and protects partial progress, a child executing an exploratory sub-task redirects toward hypothesis expansion, and a child executing a generative sub-task transitions to lower-commitment creative exploration. The downward propagation of suspension ensures that the suspension of an executive agent does not leave subordinate agents executing unsupervised tasks while their authorizing parent has itself ceased to act.

Peer-to-Peer Coordination Through a Shared Confidence Context

In peer-to-peer coordination scenarios, where agents collaborate without a strict parent-child hierarchy, confidence propagation operates through a shared confidence context. Each participating agent publishes its current confidence value to the shared context, and each agent incorporates the confidence values of its collaborators into its own confidence computation. Rather than confidence flowing along a delegation edge, it is exchanged through a common structure visible to all collaborators.

A collaborative task in which multiple agents contribute to a shared outcome has its aggregate confidence bounded by the confidence of the least-confident participating agent. This bound ensures that the collaborative effort does not proceed at a pace or commitment level that exceeds the assessed sufficiency of any participant. A single collaborator whose confidence has collapsed is sufficient to constrain the whole, which prevents a confident majority from carrying a collaboration past the readiness of a struggling member.

Confidence Contagion in Delegation Chains

The child agent's confidence value is periodically transmitted to the parent agent through the delegation communication channel. If the child agent's confidence drops below a delegation confidence threshold, which may be the same as or different from the child's own execution authorization threshold, the parent agent's confidence computation subsystem receives a delegation-adverse signal proportional to the magnitude of the child's confidence drop, scaled by a delegation importance weight that reflects the significance of the delegated subtask to the parent's overall task. This delegation-adverse signal reduces the parent's confidence.

The reduced parent confidence does not merely register passively. It may cause the parent to recall the delegation, to reassign the subtask to an alternative delegate, or to suspend its own execution pending resolution of the delegation confidence gap. Conversely, when the child agent's confidence is high and stable, the parent receives a delegation-positive signal that supports the parent's confidence maintenance. The result is a confidence-aware delegation network in which confidence information flows bidirectionally through delegation hierarchies, enabling parent agents to detect and respond to downstream confidence degradation before it results in execution failure.

Composition With Single-Agent Confidence Governance

Multi-agent propagation does not replace the single-agent confidence governor; it feeds it. The confidence values that propagate, whether upward from a child, downward as a suspension signal, or sideways through a shared context, enter the same confidence evaluation function that the receiving agent applies to its own agent state and task state inputs. A propagated confidence signal is one input among capability sufficiency, resource availability, integrity state, affective state, and the other dimensions the function already weighs. The receiving agent's confidence governor then gates execution exactly as it would for a solitary agent.

Because every mutation to the confidence field is recorded in the agent's lineage, the effects of propagation are auditable. Governance infrastructure can trace how a child's confidence drop reduced a parent's confidence, how that reduction crossed the parent's authorization threshold, and how the parent's suspension then propagated downward to its remaining children. The propagation leaves a continuous record rather than an opaque assertion that an agent stopped acting.

Why Propagation Matters

Without propagation, each agent in a coordinated system would assess its own sufficiency in isolation, blind to the state of the agents it depends on or collaborates with. A parent could continue committing to a task while a child responsible for a critical sub-task had quietly lost confidence, and a collaboration could advance at the pace of its most confident member while a struggling participant fell behind unaccounted for. Propagation closes these gaps by making confidence a shared, directional property of the executive graph rather than a private property of each agent.

The directionality is the point. Child confidence rises to inform parents who depend on it, parent suspension descends to halt children who answer to it, and peer confidence is pooled and bounded by its weakest contributor. Each direction enforces a structural relationship that already exists in the coordination topology, so that the system's collective willingness to act tracks the actual sufficiency of the agents doing the work.

Disclosure Scope

Multi-agent confidence propagation and coordination, comprising the upward propagation of child agent confidence into a parent agent's confidence computation weighted by sub-task criticality, the downward propagation of parent execution suspension to child agents that then suspend according to their task class, the peer-to-peer shared confidence context in which each agent publishes its confidence and the aggregate collaborative confidence is bounded by the least-confident participant, and the confidence contagion in delegation chains in which a child's drop below a delegation confidence threshold produces a delegation-adverse signal scaled by a delegation importance weight that may cause the parent to recall, reassign, or suspend, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). This article describes that disclosed mechanism. The scope extends to executive graphs of arbitrary depth and to coordination topologies combining hierarchical delegation and peer collaboration, provided confidence propagates directionally and the receiving agent incorporates the propagated value into the same confidence evaluation function that gates its own execution.