Written by Joel Ayuba, Lead Technical Analyst. AI architecture deconstruction based on academic game AI literature, engine profiling, and reverse-engineered combat state machines.
When you square off against Malenia, Blade of Miquella in Elden Ring, or confront the lethal precision of Isshin the Sword Saint in Sekiro, the battle feels deeply personal. The adversary seems to study your habits, feigning vulnerability only to punish your healing flask, dodging your spells with uncanny reflexes, and escalating aggression the instant your stamina depletes.
To the player, it looks like sentient martial mastery.
To the game engine, it is a complex execution of Hierarchical Behavior Trees (HBTs), Utility AI scoring functions, Blackboard memory state registers, and token-based aggression management.
How do modern game designers craft boss AI that pushes players to their absolute limit without feeling cheap, broken, or mathematically unfair?
---
1. Beyond Finite State Machines: The Rise of Behavior Trees
In the early eras of 3D gaming, enemy AI was governed by simple Finite State Machines (FSMs):
- Idle -> Patrol -> Chase -> Attack -> Flee.
While FSMs worked for straightforward grunts, they broke down completely in complex boss encounters. Adding twenty different attack strings, evasive retreat options, posture breaks, and phase transitions created an explosive web of state transitions that was impossible to debug. If a boss was interrupted mid-swing by a player stun, the state machine frequently locked up or glitched into loops.
Modern game engines solve this with Behavior Trees:
- Selectors (?): Execute children in priority order until one succeeds.
- Sequences (->): Execute children in lockstep; if any child fails, the whole branch aborts.
- Decorators / Conditionals: Gate branches based on parameters (e.g., *"Is player distance < 4 meters?"• or "Is boss health < 50%?").
---
2. Blackboard Architecture: The Boss’s Dynamic Memory
A Behavior Tree on its own has no memory; it simply evaluates nodes. To make decisions, the tree reads and writes to a centralized memory structure called the Blackboard.
In a game like Elden Ring or God of War Ragnarök, the boss's Blackboard continuously tracks:
- Vector Variables: Distance to player, angle offset (is player flanking behind?), line of sight, distance to arena boundaries (to avoid getting wedged into corners).
- Player State Flags: Is player currently in active attack recovery frames? Is player initiating an Estus / flask healing animation? Is player holding a shield block?
- Boss Internal Economics: Posture / stagger meter, remaining poise health, current combat phase (Phase 1 vs Phase 2), and individual skill cooldown timers.
Input Reading vs State Heuristics
Gamers often accuse FromSoftware bosses of "input reading"—the unfair practice where the AI detects the exact hardware button press of your controller before the player's character even moves.
What actually occurs in the code is more nuanced:
1. When you press the heal button, your character transitions into an Item_Use_Startup animation frame.
2. In the very first animation frame (millisecond 16), the game engine flips the Blackboard boolean: Player_Is_Healing = TRUE.
3. In the boss's Behavior Tree, a high-priority Selector evaluates this flag on its next tick:
IF Player_Is_Healing == TRUE AND Distance > 6m THEN Execute Lunging_Thrust_Attack
4. Because the boss AI evaluates world state every 16 to 33 milliseconds, it immediately triggers the thrust, punishing you before your potion reaches your lips.
It is not reading your physical controller wires; it is reacting to character animation startup states with near-instantaneous algorithmic reflexes.
---
3. Utility AI: The Mathematical Weighting of Decisions
While Behavior Trees excel at reactive prioritization (stunning, dodging, phase swaps), they can produce repetitive attack patterns if the same condition always triggers the same response.
To make combat organic and unpredictable, modern engines integrate Utility AI.
Instead of binary True/False branches, every potential attack is assigned a dynamic score calculated through response curves evaluating distance, player health, stamina pacing, and random entropy.
| Potential Boss Action | Close Range (< 3m) | Mid Range (3m - 8m) | Long Range (> 8m) | Player at Low Health |
|---|---|---|---|---|
| Quick Overhead Slash | Score: 92 (High) | Score: 30 (Misses) | Score: 0 (Invalid) | Score: 75 |
| Whirlwind Gap-Closer | Score: 15 (Too close) | Score: 88 (Optimal) | Score: 45 (Out of range) | Score: 95 (Execute!) |
| Defensive Backstep | Score: 60 (If high poise) | Score: 40 | Score: 10 | Score: 5 (Stay aggressive) |
| Delayed Heavy Sweep | Score: 85 (Roll-Catch) | Score: 20 | Score: 0 | Score: 50 |
By calculating utility curves on the fly, the boss selects actions that make tactical sense without becoming completely predictable.
---
4. The Illusion of Fairness: Aggression Tokens and Director Systems
If a boss AI utilized 100% of its computational capability at all times, no human player could ever defeat it. A computer can calculate perfect frame-trap parries, attack continuously without resting, and position itself with sub-millimeter precision.
Great game design is not about creating unbeatable algorithms; it is about engineering a thrilling, theatrical duel that allows human triumph through pattern recognition.
Aggression Token Architecture
In multi-enemy encounters (such as battling a boss alongside summoned minions, or fighting duo bosses like Ornstein & Smough or the Godskin Duo), engines use Aggression Tokens:
1. The global combat manager holds a limited supply of "Attacking Tokens" (often only 1 or 2 tokens).
2. An enemy cannot initiate an active offensive combo string unless it claims an available Attack Token from the manager.
3. While Enemy A holds the token and executes an attack, Enemy B is denied a token. Enemy B’s Behavior Tree falls back to low-threat pacing nodes: circling the player, maintaining spacing, and taunting.
4. When Enemy A finishes its recovery frames, it surrenders the token, allowing Enemy B to claim it and engage.
This invisible token system prevents players from being continuously stun-locked by simultaneous, overlapping attacks from blind angles, creating clean windows of opportunity.