System selector and architecture

Layer the problem

World collision -> NavMesh/path graph -> Path following/movement
Stimuli/events -> Perception/memory -> Decision architecture
World candidates -> EQS/direct query -> chosen goal
Interaction opportunities -> Smart Object claim -> interactor behavior
Nearby moving agents -> RVO or Detour Crowd -> local velocity adjustment

A bug belongs to one layer until evidence shows otherwise. Do not “fix AI” by changing several layers together.

Decision selector

Need	Default
One or two simple modes	component/enum/timers/events
Reactive priorities, conditions, abort-to-higher-priority behavior	Behavior Tree + Blackboard
Explicit hierarchical state, transitions, sequences, shared task data	StateTree
Large data-oriented crowds	Evaluate Mass/StateTree architecture separately; ordinary Actor AI may not scale

Behavior Tree and StateTree can coexist at a system boundary, but do not duplicate the same state machine in both. Assign one owner and exchange compact state/events.

Query selector

• Direct known actor/location: use the known reference or subsystem lookup.
• Reachable random point: navigation query.
• “Best cover/attack/resource point” from multiple constraints: EQS.
• Reusable reservable activity: Smart Object query/claim.
• Long-range route: NavMesh pathfinding.
• Short-range agent avoidance: RVO or Detour Crowd.

Authority model

In networked games:

• run authoritative AI controller, decisions, path requests, Smart Object claims, damage, and gameplay state on the server;
• replicate the Pawn and observable gameplay state/animation inputs as designed;
• keep client-only cosmetic prediction/presentation non-authoritative;
• do not replicate entire Blackboards or internal decision graphs by default.

Route replication details to $unreal-replication.

Data ownership

• Controller owns decision/path-following concerns tied to possession.
• Pawn/Character owns body, movement component, collision, animation-facing state.
• Blackboard holds Behavior Tree working memory, not universal game truth.
• StateTree context/external data supplies required owner/world data; Parameters configure instances.
• Subsystems/game state own durable shared facts.
• Perception stores observations; gameplay systems decide what those observations mean.

Update policy

Prefer event-driven updates for target acquired/lost, damage, state changes, path completion, Smart Object invalidation, and Blackboard changes. Use bounded timers/services for genuinely periodic sensing or scoring. Stagger expensive work across agents and treat reaction delay as a design variable, not merely a performance defect.