Headline numbers

Per-cycle wall time for one full replan pass across N agents, plus per-tick dispatcher cost at the same N.

At 2000 agents, replan + dispatcher together stay well under 1 ms per frame on a single thread. Past the 2000 budget the scaling stays roughly linear: 4000 agents land at 0.87 ms, in line with an O(N) loop.

Reference machine

Benchmark catalogue

Automation tests live under the IntentForge.Performance.* namespace.

Source: Plugins/IntentForge/Source/IntentForgeCore/Private/Tests/PerfBenchmarks_*.cpp.

Reproduce locally

"C:\Program Files\Epic Games\UE_5.7\Engine\Build\BatchFiles\Build.bat" \
  CrucibleEditor Win64 Development -Project="<path>\Crucible.uproject" -WaitMutex

"C:\Program Files\Epic Games\UE_5.7\Engine\Binaries\Win64\UnrealEditor-Cmd.exe" \
  "<path>\Crucible.uproject" \
  -ExecCmds="Automation RunTests IntentForge.Performance; Quit" \
  -unattended -nopause -nullrhi

A full sweep takes 10–30 seconds on a developer machine.

Anti-flap stack overhead

The anti-flap stack adds runtime cost on every replan and on every scalar fact write. The AntiFlapSelectionCost benchmark isolates the momentum dimension by running the same archetype with GoalMomentumBonus = 0, then with the default 0.15.

At N=500, 1000 iterations per config, 3 runs:

The delta sits inside measurement noise. The momentum bonus is effectively free — one multiplication per candidate goal during ranking, lost in the rest of the selection path.

EMA filter and Schmitt latch costs live inside FWorldState::SetScalar and only fire when the schema configures them. The WorldStateSetScalarCost micro-benchmark covers those directly across four configs (baseline / filter-only / latch-only / both) at K=100 batched calls per timing window.

Memory footprint

AgentMemoryFootprint snapshots process memory before and after spawning N components on a transient world, force-GCs, snapshots again, and divides the delta by N.

The N=100 row falls to zero because process memory snapshots vary by tens of KB between quiescent reads. N=1000 gives the stable estimate: roughly 220 bytes per agent including the component, the live FWorldState, and two history ring buffers at default depths.

Plan on "under 1 KB per agent for the framework data" plus whatever your sensors and executors carry.

Sensor cost

Per-Core-sensor cost for one Sample() call, batched K=100 per timing window across 5000 windows.

LineOfSightToActor does an actual line trace. The empty-scene case above is best case — expect 2–5× in a dense scene with collision geometry.

Perception sensors (AIPerception, BlackboardBool/Float, EnvQuery) are excluded by design: their cost is dominated by the AI subsystems they bridge into, not the sensor pattern itself.

Methodology

• Warmup: 10 iterations per config before timing.
• Repeats: 1000 iterations per config (100 at N ≥ 2000). Multi-agent and dispatcher benchmarks do 3 runs and report median p99.
• Percentiles: nearest-rank, no interpolation.
• Lead statistic: p99. A mean that hides a 10× spike every 100 frames is a hung-frame problem.
• GC: full purge between runs.
• Timer: FPlatformTime::Seconds() (Windows QPC ≈ 100 ns resolution). Sub-microsecond benchmarks batch K=100 calls per window to clear the noise floor.

Caveats

• Single-thread. No async path. The numbers are the full game-thread cost.
• No movement / animation / nav in the test worlds. Real games carry per-actor cost outside the framework.
• ManyAgentReplan measures per-cycle wall time, not per-plan compute. Coalescing and plan reuse mean the per-cycle work is less than "plan from scratch per agent" — which matches realistic production.
• AgentMemoryFootprint uses process-global stats. Noisy at the tens-of-KB level — treat as order-of-magnitude.
• CPU variance is real. Consumer CPUs differ by 2–3× single-threaded. Your numbers will differ; the shape (linear, bounded tail) is what to look for.

Community runs

Re-run the benchmarks on your own hardware and add a row via PR. If you drop the CSV alongside, the project keeps a per-machine record.