Game AI rarely has one exact correct output. Verify invariants and acceptable outcomes, not a brittle sequence of internal nodes.
Build small fixtures
Create minimal maps for perception, doors, cover, path links and squads. Control start state and random seed.
Test properties
- Agents never enter non-walkable space.
- Lost perception clears direct tracking.
- Cancelled tasks release reservations.
- Decision updates respect budget.
- A fallback runs when planning fails.
Add soak tests
Run crowds for long periods, change routes and remove targets mid-action. Monitor stuck time, replans, allocations and queue growth.
Keep playtests
Automation finds regressions; humans judge fun and readability. Convert confirmed bugs into focused cases and preserve traces from failures.