Forecasting as Coordination Primitive
by Nick Clark | Published March 27, 2026
Multi-agent coordination is structured around the forecast as the elementary primitive. Agents publish bounded forecasts of their intended near-future trajectory and align their commitments around the regions in which their forecasts overlap. Convergence under bounded forecast publication is provable; the protocol does not require a centralized scheduler, a designated leader, or a consensus round.
Mechanism
Each agent maintains an internal forecasting engine that constructs a planning graph over the near-future evolution of its semantic state. From this graph, the agent extracts a forecast object consisting of a typed sequence of intended transitions, each annotated with a confidence value, a temporal bound, and a contingency set indicating the conditions under which the forecast remains valid. The forecast is bounded in horizon, in branching factor, and in confidence; it does not represent a commitment, only a publishable intention.
Forecasts are published to a shared coordination channel. Other agents subscribed to the channel receive the forecast and ingest it into their own forecasting engines as exogenous structure. An ingested forecast is not committed to the receiving agent's verified state; it is incorporated into the receiving agent's planning graph as a typed external branch. The receiving agent's planner can then evaluate alignment between its own intended trajectory and the published trajectory, scoring overlap regions where commitments are mutually reinforcing and divergence regions where commitments would collide.
Coordination emerges from the alignment scoring. Each agent independently selects, from its own planning graph, the trajectory that maximizes alignment with overlapping published forecasts subject to its local objective and integrity constraints. Because every agent applies the same deterministic alignment evaluation to the same set of published forecasts, the selected trajectories converge on the overlap regions without requiring negotiation, voting, or external arbitration. The coordination primitive is therefore the forecast itself, and the coordination operation is the local maximization over published material.
Convergence under bounded forecast publication is established by treating each round of publication and re-selection as a contraction over the joint trajectory space. Bounding the forecast horizon and confidence ensures that the contraction has a fixed point; bounding the branching factor ensures the fixed point is reachable in a finite number of rounds. The protocol terminates when no agent's locally optimal trajectory changes between successive rounds, at which point the published forecasts collectively describe a coordinated plan.
Operating Parameters
The forecast horizon parameter bounds the temporal extent of any published forecast. The branching-factor parameter bounds the number of alternative trajectories carried in a single forecast object. The confidence floor parameter excludes branches below a specified confidence from publication, ensuring that low-quality speculation does not pollute the coordination channel. The republication interval parameter governs how frequently an agent republishes a refreshed forecast, trading responsiveness for channel load.
The alignment-scoring function exposes weights for overlap reward, divergence penalty, and contingency-set compatibility. These weights are declared in the policy reference and are identical across cooperating agents in a given coordination context. Heterogeneous weights are permitted across coordination contexts, allowing the same agents to coordinate under different regimes for different objectives, but within a single context the parameters are fixed for the duration of the convergence run.
Alternative Embodiments
In a first embodiment, the coordination channel is a publish-subscribe broker delivering forecasts in soft real time to a fixed cohort of cooperating agents. In a second embodiment, the channel is a gossip overlay in which forecasts propagate through pairwise exchanges among agents with no global broker. In a third embodiment, the channel is a shared append-only log to which forecasts are written and from which agents read at their own pace, decoupling publication from ingestion. The convergence argument holds in each topology, modulo the propagation latency of the channel; only the wall-clock time to convergence differs.
The cohort composition may be static, as in a fixed team of cooperating agents, or dynamic, with agents joining and leaving across rounds. Joining agents publish an initial forecast and ingest the current set; departing agents publish a terminating forecast that releases the regions they previously occupied. The protocol does not require global membership knowledge; each agent maintains only the set of forecasts it has received and not yet superseded.
Heterogeneous proposer embodiments are also supported. An agent backed by a transformer-based planner, an agent backed by a search-based planner, and an agent backed by a hand-authored policy can participate in the same coordination run, provided each emits forecast objects in the canonical schema. The coordination primitive is defined at the schema level, not at the planner level.
Composition
The forecasting engine composes with the inference-control layer such that every transition considered for inclusion in a published forecast passes through the same admissibility evaluation that governs committed transitions. Forecasts therefore cannot include trajectories that would be rejected upon execution, and the coordination channel is protected from being populated with inadmissible speculation. The trust slope vector maintained by the control surface feeds into the confidence floor used by the forecasting engine, so that an agent experiencing a degraded admissibility trajectory automatically narrows the forecasts it is willing to publish.
Lineage records produced during forecast construction, publication, ingestion, alignment scoring, and trajectory selection are preserved on the same lineage substrate as the agent's committed transitions. A coordinated decision is therefore reconstructible end-to-end: from the forecasts each agent published, through the alignment scores each agent computed, to the trajectory each agent selected and committed.
Implementation Considerations
Forecast publication is implemented as a serialization of the planning graph subset selected for external visibility. Internal branches retained for local exploration but not yet ready for publication remain in the agent's planning memory and are excluded from the published object. The serialization is canonical and deterministic: the same internal graph subset produces byte-equal forecast objects across runs, allowing receiving agents to detect republished forecasts and deduplicate them in their own ingestion logic.
Ingestion is bounded by a per-agent budget. An agent receiving forecasts from a large cohort cannot incorporate every forecast into its planning graph without exceeding its planning budget. The ingestion logic therefore prioritizes forecasts by overlap with the receiving agent's current intended trajectory, by recency, and by publisher trust. Lower-priority forecasts are retained at lower fidelity or discarded; the discard decision is recorded in lineage.
Termination detection in a distributed setting is a per-agent local property. Each agent observes that its own locally optimal trajectory is unchanged across two successive rounds and marks itself as locally converged. Local convergence does not require global agreement on termination; an agent that observes inbound forecasts continuing to refine its own selection simply remains active. The protocol does not require a global termination round, and there is no coordinator responsible for declaring the run complete.
Channel latency interacts with the republication interval. A republication interval shorter than the channel propagation time produces forecast objects that are superseded before they are ingested, wasting channel capacity. A republication interval longer than the underlying task dynamics produces forecasts that are stale by the time they are evaluated. The interval is therefore tuned per deployment to the slower of the channel and the task; in practice this is a parameter exposed to operators rather than a structural property of the protocol.
Prior Art Distinction
Conventional multi-agent coordination relies on either a centralized scheduler that assigns roles and trajectories from a global vantage, or a consensus protocol in which agents exchange proposals and votes until a quorum is reached. Centralized schedulers introduce a single point of failure and require global state. Consensus protocols incur quadratic message complexity in the cohort size and require explicit termination conditions decoupled from the underlying task. In both cases, the coordination primitive is a control message rather than a model of intended behavior.
The distinction of the present mechanism is that the coordination primitive is the forecast itself: a bounded model of intended near-future behavior that other agents can ingest and evaluate against their own forecasts. There is no scheduler and no quorum. Convergence arises from local optimization against shared published material, and the protocol's termination is a property of the bounded forecast space rather than an externally imposed stopping rule.
Convergence Properties
The convergence argument rests on three structural facts. First, the bounded forecast horizon ensures that the joint trajectory space is finite for any fixed cohort and any fixed planning budget. Second, the bounded branching factor ensures that the number of trajectories any agent can publish in a single round is finite. Third, the deterministic alignment-scoring function ensures that, for fixed published material, each agent's locally optimal trajectory is unique up to declared tie-breaking.
Under these conditions, the round-by-round update is a function from the joint trajectory space to itself with a finite range. By a standard fixed-point argument, the update has at least one fixed point, and any orbit either reaches a fixed point in finitely many rounds or enters a cycle. The cycle case is precluded by the monotonicity of the alignment score under progressively more compatible forecasts: once an agent's selection moves toward an overlap region, subsequent rounds cannot reduce the achieved alignment without violating the agent's local objective. The protocol therefore terminates at a fixed point.
The fixed point is not necessarily globally optimal in the sense of jointly maximizing cross-agent alignment. It is locally optimal in that no agent can unilaterally improve its alignment by changing trajectory, holding the published forecasts of its peers fixed. This is the appropriate convergence criterion for a decentralized protocol without a global objective: the protocol delivers a coherent joint plan reachable from local optimization, not a centrally-optimal assignment.
Robustness of the convergence under partial failure is a structural consequence of the local-only contract. An agent that fails to publish in a given round contributes no new forecasts, but the previously published forecasts remain available to its peers up to their declared expiration. Peers continue to converge against the available material. An agent that publishes corrupted forecasts is detected by the canonical schema validation at the ingestion boundary; invalid forecasts are rejected before alignment scoring and the publishing agent is recorded in lineage as having produced a malformed object. The protocol does not stall on a single faulty publisher because no publisher is on the critical path of any peer's progression.
Bounds on time to convergence are deployment-dependent but structurally finite. In the worst-case configuration, the number of rounds required is bounded by the diameter of the joint trajectory space under the alignment-scoring relation. In typical configurations, convergence completes in a small number of rounds because most agents' locally optimal trajectories stabilize quickly once overlap regions become visible.
Disclosure Scope
Disclosure encompasses the canonical forecast schema; the publication and ingestion semantics on the coordination channel; the alignment-scoring function and its declared weights; the bounded operating parameters governing horizon, branching, confidence, and republication; the convergence argument under bounded publication; and the lineage schema recording each round of the protocol. The disclosure extends to all channel topologies preserving the publication and ingestion semantics, all cohort compositions consistent with local-only membership knowledge, and all heterogeneous proposer combinations satisfying the canonical forecast schema.