Game AI Behaviour and Navigation
Build deterministic, debuggable and budgeted runtime agents whose decisions, movement, perception and interaction remain legible under worst-case gameplay conditions.
Prerequisites
Load world-class-engineering, gameplay-systems-engineering, level-world-and-content-production, the selected engine skill, performance/testing skills, and the game's accessibility, difficulty and narrative contracts.
Use When
- Defining agent goals, states, decisions, behaviour trees/state machines/utility systems, blackboards, perception, navigation, avoidance, crowds or Smart-Object-style reservations.
- Debugging stuck, oscillating, unfair, expensive, nondeterministic or unreadable NPC behaviour.
- Specifying AI architecture, data contracts, budgets, telemetry, authoring tools and validation scenarios.
Do Not Use When
- The task is generative AI, model training or an online LLM feature.
- The task is animation rigging; use
game-3d-asset-pipelineand animation systems. - The requested API is engine-specific but engine/version is unknown; freeze a versioned spike first.
Required Inputs
| Input | Required | Missing-input response |
|---|---|---|
| Agent fantasy, capabilities, fairness/readability rules and difficulty envelope | yes | Stop architecture and return gameplay decisions required. |
| World topology, traversal types, collision, interactions and animation contract | yes | Limit work to an abstract state model. |
| Pinned engine/version and feature maturity policy | yes for implementation | Mark APIs and performance not assessed. |
| Simultaneous-agent envelope, frame/memory/network budgets and target hardware | yes | Block scale claims and crowd architecture. |
| Determinism, save/load, replay, authority and multiplayer constraints | conditional | Flag state ownership and reproduction gaps. |
Workflow
- Define each agent archetype by player-observable goals, verbs, senses, knowledge limits, reaction windows, failure modes and fairness contract.
- Separate sensing, memory/world state, decision, action execution, navigation, animation and feedback. Keep engine controllers/adapters outside reusable domain logic.
- Choose the simplest decision model that explains the behaviour: finite state for bounded modes, behaviour tree for reactive hierarchy, utility for competing scored choices, planner for expensive compositional goals, or hybrid with explicit boundaries.
- Give all shared state typed keys, ownership, defaults, lifetime and update/event policy. Prefer event-driven reevaluation; bound polling rates.
- Specify navigation agents, mesh generation/update, areas/costs/filters, links, path failure, partial paths, moving targets, avoidance, off-mesh traversal and recovery.
- Treat interactable locations as capacity-controlled resources: eligibility, reservation, timeout, cancellation, completion and orphan cleanup are explicit.
- Integrate animation and gameplay with named events and bounded waits. Define root-motion, turning, hit reaction, interruption and fallback contracts.
- Instrument decision transitions, perception changes, path requests/failures, stuck recovery, reservation lifecycle and per-archetype cost without shipping noisy debug output.
- Test single-agent correctness, adversarial geometry, simultaneous worst case, save/load, network authority, difficulty variants and target hardware. Retain seeds, scenario, build and trace.
- Stop on unbounded search/polling, missing recovery, unstable experimental dependency, unfair information access, unreachable state or budget failure.
- Load
references/behaviour-telemetry-and-tuning.mdbefore tuning behaviour. Retain the exact build, seed, configuration, scenario, state trace, path failures, and guardrail result for every experiment; never promote learned or probabilistic behaviour without a deterministic fallback and rollback path.
Outputs
| Artefact | Acceptance condition |
|---|---|
| AI behaviour contract and state/key registry | States, transitions, ownership, defaults and failure recovery are deterministic. |
| Navigation/interaction contract | Agents, costs, links, reservations, invalidation and fallback are explicit. |
| Engine adapter and feature-status decision | Version, plugin/feature maturity, fallback and upgrade evidence are recorded. |
| AI verification and profiling report | Named scenarios prove correctness, legibility and worst-case cost. |
Evidence Produced
| Category | Artifact | Format | Example |
|---|---|---|---|
| Correctness | Transition, perception, path, reservation, save/load and recovery traces. | Trace log with build, seed, scenario and state transitions | docs/game-ai/navigation-traces.jsonl |
| Performance | Per-agent and simultaneous worst-case CPU, memory and path-query evidence. | Profiling report with scenario, target hardware and measured costs | docs/game-ai/navigation-profile.md |
| UX quality | Reaction, telegraph, knowledge-boundary and difficulty playtests. | Playtest report with fairness/readability observations and findings | docs/game-ai/fairness-playtests.md |
| Release evidence | Engine/plugin version, feature status, known failure states and regression suite. | Release record with version identity, limitations and regression results | docs/game-ai/release-evidence.md |
Capability and Degraded Mode
Changing navigation generation, runtime authority or experimental plugins requires implementation authority and rollback. Without the pinned engine and representative build, provide contracts and test scenarios only; mark API fit, feature stability and performance not assessed.
Decision Rules
| Condition | Action |
|---|---|
| Behaviour can be expressed with a small explicit state machine | Do not introduce a planner or deep tree. |
| Agent reads information the player could not reasonably expose | Add a perception/memory path or label the intentional difficulty rule. |
| Decision checks run every frame without need | Convert to events, services at bounded intervals, or scheduled batches. |
| Path or interaction can fail | Specify timeout, retry/backoff, alternate goal and safe idle—not an infinite wait. |
| Engine feature is beta/experimental | Record status in the pinned release, isolate behind an adapter and define fallback. |
Quality Standards
- Name behaviours by intent; keep tasks/actions small, reusable and independently testable.
- Do not mutate decision-node configuration at runtime; keep changing state in typed runtime data.
- AI difficulty changes decision quality, timing, resources or coordination explicitly—not hidden omniscience.
- Measure worst cases with actual agents, geometry, animation and target hardware.
- Record time and memory complexity, update frequency, worst-case path/search cost, and invariants for critical behaviours. Test a bounded degraded path before standardising a Kaizen change.
Anti-Patterns
- Giant behaviour tree as a visual script. Fix: split domain decisions, reusable actions and engine adapters.
- Polling everything every frame. Fix: event-driven updates and budgeted schedules.
- Navigation success treated as guaranteed. Fix: handle invalid targets, partial paths, topology change and stuck recovery.
- Reservation without cancellation. Fix: define ownership, timeout and cleanup on death/unload/disconnect.
- Debugging by animation alone. Fix: capture state, reason, key changes, path status and timestamps.
- Tutorial agent counts treated as product budgets. Fix: measure project worst cases.
Read Next
- The selected Unity/Unreal/Godot skill for concrete implementation.
level-world-and-content-productionfor nav-ready spaces and encounter topology.game-testing-polishand performance skills for scenario and target-device evidence.
