← Back to OverLift
OVERLIFT NPC AGENT FABRIC · GOVERNED GAME INTELLIGENCE

Autonomous characters.
Deterministic control.

OverLift NPC Agent Fabric gives game characters their own identity, perception, bounded memory, goals, personality, permissions, and decision loop—without giving a language model control of physics, quests, inventory, score, or official world state.

The objective is not endless AI dialogue. It is behavior that notices what the player actually does, remembers meaningful patterns, coordinates with other characters, changes tactics, and knows when silence is better than another recycled line. Repetition is treated as a defect signal, and the player’s behavior trail can remain inside the game instead of becoming a cloud transcript.

Open Neon Drift Explore OverDrift CITADEL Inspect the architecture
CHOOSE YOUR LENS

One character system. Four reasons to buy or build it.

Start with the product value that matters to you, then inspect the evidence behind it.

NO MORE REPETITIVE INTERACTIONS

A character should respond to the moment—not restart a bark loop.

The Fabric treats dialogue, movement, timing, refusal, and silence as one governed behavior problem. The player experiences adaptation because the character changes what it does, not because a model produces more words.

REMEMBER CONTEXT

Meaningful memory, not an endless transcript.

Each agent keeps bounded, typed state about patterns that matter to gameplay.

  1. 01
    Track player tendencies

    Routes, timing, risk tolerance, preferred attacks, escape habits, and repeated feints can become explicit evidence.

  2. 02
    Separate short and durable memory

    Immediate tactical context expires quickly; proven patterns may persist under versioned limits.

  3. 03
    Forget safely

    Stale, contradicted, low-confidence, or unauthorized memory is retired instead of becoming folklore.

DETECT REPETITION

Conceptual duplicates are defects.

Different wording does not make the same interaction fresh.

  1. 01
    Compare meaning, not just strings

    A semantic service can identify when two taunts, warnings, or plans communicate the same idea.

  2. 02
    Bind repetition to context

    A line may be valid again only when the world state, speaker intent, and player history materially justify it.

  3. 03
    Prefer a new action

    When speech would repeat, the agent can feint, reposition, wait, deny a route, or remain silent.

BEHAVIOR BEFORE CHATTER

The best response may be physical.

Intelligence should change play, not cover static behavior with commentary.

  1. 01
    Adapt tactics

    Cut off a preferred escape, protect an objective, hold a chokepoint, retreat, or change formation.

  2. 02
    Coordinate visibly

    A director can propose a plan while an embodied NPC accepts, modifies, delays, or rejects it.

  3. 03
    Let consequences speak

    The player should recognize that the character learned before the interface explains why.

PACE THE EXPERIENCE

Silence protects character and tension.

A pacing governor decides whether communication is timely, useful, and still true.

  1. 01
    Cooldown by concept and speaker

    Repeated ideas, overlapping voices, and low-value commentary are suppressed.

  2. 02
    Cancel stale delivery

    A line or directive that no longer fits the authoritative event is withdrawn before it reaches the player.

  3. 03
    Abstain without failure

    The system may deliberately say nothing when no permitted action improves the moment.

NEON DRIFT CASE STUDY

Two minds. Two jobs. One authoritative arena.

Zarvox is the intelligence above the arena: it observes, judges, speaks, and proposes directives. VANTA-9 is the singular red hunter-cycle inside the arena: it perceives, predicts, hunts, and owns a separate tactical decision context.

The player-facing rule Zarvox is the mind above the arena. VANTA-9 is the red machine inside it.
The authority rule Neither agent may directly mutate movement, collisions, score, timing, inventory, or world state. The compiler and world validate every consequential action.
ARENA INTELLIGENCEZARVOX // VANTA-9, CLOSE THE EAST CHANNEL.

Zarvox can recognize a player pattern and propose a bounded directive without impersonating the hunter.

EMBODIED AGENTVANTA-9 // REJECTED. COLLISION CERTAIN.

VANTA-9 can reject an unsafe directive and preserve its own tactical identity instead of obeying blindly.

ADAPTIVE PROFILESHUNTER · PREDATOR · SENTINEL · ADAPTIVE

Different profiles reshape ranking and temperament over one legal action set; permissions and world rules do not change.

CURRENT REFERENCE PROOFND-OL01…06 + ND-NPC01…04

The implemented reference spans governed semantics, Q-Lens, compilation, replay, VANTA-9 authority, coordination, profile comparison, failure injection, and an engine-neutral proof boundary.

Neon Drift gameplay showing the Zarvox voice ticker, VANTA-9 hunting the player, power cores, and the deterministic neon arena
Two-agent gameplay proofZarvox communicates above the arena while VANTA-9 acts inside it under the same authoritative world rules.
ARCHITECTURE

Shared intelligence. Separate agents. Immutable world authority.

The Fabric is a control plane between perception and action. It lets many characters share efficient semantic services while keeping identity, memory, permissions, goals, and executable authority explicit.

LAYER 1

Authoritative world and perception fabric

The game publishes only what each agent is permitted to perceive.

  1. 01
    World remains sole writer

    Physics, collisions, quests, inventory, score, time, legality, and outcomes stay in the game engine.

  2. 02
    Typed observations

    Perception packets carry identity, freshness, provenance, visibility, and replay keys.

  3. 03
    Fast snapshots when needed

    SharedArrayBuffer may carry fixed-size, high-frequency world snapshots; it never becomes the semantic conversation bus.

LAYER 2

Independent agent capsules

Every NPC is a governed identity—not another prompt in one global chatbot.

  1. 01
    Identity and personality

    Voice, temperament, role, relationships, and allowed expression are versioned separately.

  2. 02
    Memory and goals

    Bounded tactical and narrative memory supports continuity without unbounded context growth.

  3. 03
    Permissions and action catalog

    An agent may propose only actions its capsule and current world state permit.

LAYER 3

Semantic and path-evaluation fabric

Meaning and consequence help rank legal choices; they do not create new authority.

  1. 01
    Shared semantic service

    One local inference service can support many bounded agents instead of loading one model per NPC. In Neon Drift, pinned MiniLM reranks only admitted candidates and helps detect conceptual repetition; deterministic behavior continues when it is unavailable.

  2. 02
    Q-Lens consequence paths

    Response, movement, timing, delivery, delay, abstention, and cancellation can be evaluated together.

  3. 03
    Localized inertia

    The field may reshape the search landscape; it may not rewrite facts, policy, provenance, permissions, deadlines, or world rules.

LAYER 4

Compiler, outcome receipts, and replay

One deterministic boundary decides what can actually reach the world.

  1. 01
    Compile one exact action

    Freshness, candidate membership, permissions, idempotency, fairness, and world constraints are revalidated.

  2. 02
    Verify the result

    The world reports what happened; the agent does not get to declare its proposal successful.

  3. 03
    Seal and replay

    Perception, memory, candidates, evidence, decision, compiled action, and outcome are bound into inspectable proof.

ENGINEERING PIPELINE

From authoritative event to replayable behavior.

The same bounded sequence works for speech, movement, tactical coordination, companion assistance, crowd intent, or simulation actors.

  1. WORLD EVENTThe engine records what actually happened, who may perceive it, and which revision is authoritative.
  2. PERCEPTIONA typed packet gives one agent only the visible, fresh, permission-bound state it needs.
  3. ADMISSIONRules reject stale events, illegal actions, duplicates, tampering, and unsupported memory before reasoning begins.
  4. EVALUATIONSemantic evidence and Q-Lens compare only the admitted candidates, including delay, refusal, and silence.
  5. COMPILEThe deterministic compiler validates one exact action against current world state and fairness constraints.
  6. OUTCOMEThe world executes or rejects the action, then emits an outcome receipt that updates bounded memory and supports replay.
ENGINEERING BOUNDARIES

Fast where it should be. Explicit where it must be.

High-frequency simulation data and semantic intent have different jobs. The Fabric avoids a single shared mutable bus that would blur state, speech, permissions, and execution.

SharedArrayBuffer boundary Use it selectively for fixed-size authoritative snapshots and telemetry shared with workers. The world owns writes; readers consume versioned frames.
Typed-message boundary Use versioned postMessage or MessageChannel contracts for directives, responses, dispositions, dialogue intents, voice intents, receipts, and replay.
PERCEPTION CONTRACTneondrift.npc-perception.v1

What the agent may see, when it was true, who authored it, and which world revision it describes.

AGENT MESSAGEneondrift.agent-message.v1

A typed directive or report with sender, recipient, intent, deadline, and allowed dispositions.

ACTION PROPOSALneondrift.action-proposal.v1

One legal candidate with evidence, cost, timing, assumptions, permissions, and expected outcome.

COMPILED ACTIONneondrift.compiled-action.v1

The single executable operation authorized after freshness, policy, state, and idempotency checks.

FAILURE QUARANTINEreplay divergence → reject + retain baseline

Malformed evidence, stale directives, replay divergence, or failed quality guards cannot leak into world authority.

PRIVACY ADVANTAGE

Player behavior can stay in the game—not become someone else’s dataset.

The Fabric can learn useful local patterns without requiring a raw play-session transcript, one hosted model per character, or cloud authority over the world.

LOCAL PERCEPTION

Agents receive typed world packets, not unrestricted telemetry.

Each character sees only the fresh, permitted state required for its current role and action catalog.

BOUNDED MEMORY

Keep useful patterns without retaining an endless player transcript.

Durable memory is versioned, limited, reversible, and retired when it becomes stale, contradicted, or unauthorized.

SHARED LOCAL SEMANTICS

Many characters can share one local service.

Concept comparison and repetition detection do not require a separate remote model—and separate data export—for every NPC.

PROVIDER-OPTIONAL AUTHORITY

The game remains playable, exact, and replayable without hosted inference.

If an external provider is used, it receives a deliberately minimized packet; identity, permissions, world state, compilation, and outcome receipts stay local and authoritative.

SECURITY ADVANTAGE

An autonomous character gets an action catalog—not the game engine.

Agent autonomy is useful only when perception, memory, communication, generated intent, and executable world changes remain different authority levels.

CAPABILITY-LIMITED PERCEPTION

Each agent receives only its permitted view.

Role, range, freshness, occlusion, difficulty, and game rules decide which typed world facts the character may observe.

DECLARED ACTION CATALOG

The agent cannot invent an executable verb.

It may select or propose only versioned actions the host has declared, with bounded parameters, cost, deadline, and legal-state preconditions.

DETERMINISTIC COMPILER

Intent is not execution.

C17/WebAssembly authority revalidates identity, state revision, collision, physics, inventory, cooldown, policy, and idempotency before a proposal can affect the world.

REPLAYABLE FAILURE

Malformed, stale, illegal, or divergent behavior fails closed.

The runtime rejects the proposal, preserves the authoritative baseline, records the denial, and can reproduce the same decision for QA.

INTELLIGENT NPCs ACROSS GAMES

One Fabric. Many kinds of character intelligence.

The runtime is not tied to racing or taunts. The same authority pattern can govern rivals, companions, squads, crowds, directors, narrative characters, and training simulations.

RIVALS AND HUNTERS

Opponents that study the player.

Adaptive pursuit without unfair physics or hidden stat cheating.

  1. 01
    Recognize patterns

    Observe escape routes, timing windows, defensive habits, and risk choices.

  2. 02
    Change tactics

    Feint, intercept, deny space, protect an objective, or retreat.

  3. 03
    Stay fair

    Perception and action permissions prevent omniscience and impossible execution.

COMPANIONS AND SQUADS

Allies with judgment, not obedience scripts.

Characters can coordinate while preserving role, competence, and personality.

  1. 01
    Interpret intent

    Understand a player request in the context of current goals, danger, resources, and role.

  2. 02
    Negotiate safely

    Accept, modify, delay, reject, or abstain when an order conflicts with survival or policy.

  3. 03
    Explain sparingly

    Use concise communication only when it improves coordination or player trust.

CROWDS AND FACTIONS

Many identities without one model per actor.

Shared services support scalable behavior while capsules preserve local state.

  1. 01
    Shared semantic infrastructure

    One bounded service can classify events and compare intent for many agents.

  2. 02
    Independent social memory

    Each character or cohort tracks only permitted relationships, events, and attitudes.

  3. 03
    Deterministic arbitration

    Priority, space, fairness, and performance budgets resolve competing actions.

DIRECTORS AND NARRATIVE

Pacing systems that react without rewriting canon.

Dynamic presentation can remain creative while quests and facts stay exact.

  1. 01
    Shape tension

    Adjust encounters, commentary, silence, assistance, and pressure within explicit bounds.

  2. 02
    Protect story state

    Models may suggest expression; typed systems own canon, quest flags, inventory, and consequences.

  3. 03
    Replay the experience

    Designers can inspect why the director acted, compare alternatives, and reproduce a defect.

BUSINESS CASE

Governed autonomy is the product—not another model endpoint.

Studios need adaptive characters they can integrate, tune, test, ship, support, and debug. The Fabric turns intelligent behavior into a production system with boundaries, privacy, cost control, and evidence.

GAME STUDIOS

Fresher behavior with production control

Build companions, rivals, enemies, commentators, crowds, or dynamic narrative systems without putting world authority inside a generative model.

Potential value: player retention, replayability, differentiated character behavior, and lower repetitive-content fatigue.
ENGINE AND MIDDLEWARE TEAMS

A governed agent layer, not a chatbot integration

Expose engine-neutral perception, proposal, compilation, receipt, and replay contracts across different games and model providers.

Potential value: reusable adapters, shared QA tooling, provider independence, and consistent authority boundaries.
SIMULATION AND TRAINING

Adaptive actors with auditability

Use bounded characters where repeatability, legal actions, explainability, scenario control, and post-run review matter as much as realism.

Potential value: traceable decisions, controlled difficulty, reproducible scenarios, and safe experimental comparison.
COMMERCIALIZATION HYPOTHESES

What a buyer could actually purchase.

These are product hypotheses to validate through the reference demo and external pilots—not claims that a general-purpose SDK has already been adopted by studios.

Core buyer promise Autonomous characters with deterministic control, inspectable evidence, exact replay, graceful fallback, and safe rejection.
PER-TITLE RUNTIMEFabric runtime + engine adapter + Agent Lens

A bounded integration for one shipped title, including support and acceptance evidence.

ENTERPRISE / SOURCE LICENSEPortable contracts + authority boundary + replay

For engine groups that need deeper control, internal platform integration, or multiple titles.

INTEGRATION ENGAGEMENTOne vertical slice → measured pilot

Implement one rival, companion, squad, or director use case and prove quality, safety, latency, and authoring impact.

EVALUATION TOOLKITAgent Lens + Replay Lab + failure injection

Compare behavior systems, models, profiles, pacing policies, and regressions before granting gameplay authority.

MARKETING POSITION

Show behavior first. Show the evidence second.

The strongest demonstration is not a character producing a clever paragraph. It is a player recognizing that the game noticed, adapted, coordinated, protected private play context, refused something unsafe, and can prove why.

  1. 1 · PATTERNThe player repeatedly escapes through the same corridor. The arena records a meaningful tendency.
  2. 2 · ADAPTATIONVANTA-9 changes its physical plan and begins denying that route instead of repeating a taunt.
  3. 3 · DIRECTIVEZarvox proposes a coordinated intercept using a typed, expiring message.
  4. 4 · JUDGMENTVANTA-9 accepts, modifies, delays, rejects, or abstains according to its own evidence and permissions.
  5. 5 · AUTHORITYThe compiler blocks any stale, illegal, unfair, duplicated, or physically impossible action.
  6. 6 · PROOFAgent Lens and Replay Lab reveal perception, memory, candidates, decision, compiled action, and outcome.
HONEST SCOPE

What the Fabric is—and what it refuses to become.

The architecture is valuable because it preserves the control expected from gameplay systems while adding context, adaptation, personality, and coordination.

NOT A CLOUD CHATBOT

Behavior is not a conversation bolted onto a game.

Dialogue is one possible presentation channel. Movement, timing, refusal, cooperation, silence, and world-aware action matter more.

The Fabric can use local or provider-backed semantics, but gameplay must remain viable when semantic enhancement is unavailable.
NOT MODEL AUTHORITY

A confident model still cannot change the world.

Models may contribute meaning and prediction. Only typed, admitted, deterministically compiled actions can reach authoritative state.

Facts, permissions, policy, deadlines, legal actions, physics, and outcomes stay outside probabilistic control.
NOT “SENTIENT NPC” THEATER

The proof is engineering, not anthropomorphic hype.

The product claim is governed autonomy, character continuity, adaptive behavior, debugability, portability, and safe failure.

No quantum hardware claim, no magical certainty, and no promise that unlimited generated dialogue equals intelligence.
BUILD CHARACTERS THAT CAN SURPRISE PLAYERS—AND STILL PASS QA

Make adaptation visible. Keep authority exact.

Neon Drift is the reference proof. The larger opportunity is an engine-neutral Fabric for intelligent rivals, companions, squads, crowds, directors, narrative systems, and simulations that do not fall back into repetitive interactions. Read the complete OverLift Method, including the governed Semantic Search layer.

Experience Neon Drift See CITADEL in detail Discuss an intelligent NPC pilot