The Deepwoken AI Training System Is Not What The YouTube Videos Make It Look Like
I spent roughly three weeks pulling my hair out trying to get the AI companion behavior in Deepwoken to actually be useful rather than just standing around looking decorative. The community guides mostly skip over the hard parts because they don't want to admit how much trial and error is involved. Here is what actually happened when I trained mine properly. First, a clarification on terminology that most people get wrong. When Deepwoken refers to "AI training," it is not talking about a separate machine learning pipeline or any actual neural network weights you adjust externally. The system uses a behavior tree layered on top of a reward-weighting mechanic. You shape the AI through in-game actions and mantra assignments, not through a config file. This distinction matters because if you are looking for a JSON edit or a mod to force specific behavior, you will waste time looking in the wrong place. The core mechanic works like this. Your AI companion has a set of default behavioral weights for aggression, spacing, healing priority, and target selection. These weights are modified by which mantras you slot into it and how consistently it participates in combat scenarios. The more a particular behavior is reinforced through gameplay loops, the more permanent that tendency becomes in the companion's decision tree. It is not instant. You are looking at anywhere from 40 to 80 matches before you see stable behavioral patterns lock in.
Here is the practical method I ended up using. I started by giving my AI no offensive mantras for the first 15 matches. This sounds counterintuitive but it forces the companion to rely entirely on its base spacing and aggro mechanics, which reveals its default behavior tree without interference. I watched closely where it positioned itself relative to me and which enemies it engaged first. Most people skip this phase and immediately load up damage mantras, which masks the underlying tendencies you actually need to understand before you start shaping them. Once I had a baseline, I began introducing mantras in a very specific order. Aggression mantras first, then support mantras, then utility. The reason for this sequence is that aggression mantras change the target selection logic, and if you add support mantras before the companion has established who it is attacking, it will often apply healing or buffs to the wrong targets or not apply them at all during chaos. I learned this the hard way during a Tower run when my AI kept using purge on enemies instead of on me while we were being focused by the three-mantra void casters. The partner took roughly 60 percent of incoming damage that session because the behavior tree was internally conflicted. The workaround was brutal but effective. I quit the match, reset the companion's mantra loadout to empty, ran three solo practice matches in the tutorial area with just basic attacks, then reapplied the mantras one at a time between matches. This gave the behavior tree a clean slate to anchor onto each new ability. After that process, the companion's decision making stabilized noticeably. It took about six additional matches for the new patterns to fully commit, but after that it held.
Another thing nobody mentions is the terrain interaction layer. Deepwoken's AI has a separate evaluation path for environmental hazards and elevation changes. If you train your companion in flat arenas like the starter zones, it will develop poor spatial awareness for vertical combat. I noticed my AI repeatedly walking off ledges during the Eclipsed Manor runs because it had never learned to evaluate drop damage during training. I had to deliberately train it in multi-level maps for at least 20 matches before the hazard avoidance weights adjusted. This is a slow process and there is no shortcut for it other than repeated exposure to complex terrain. There are also some hard limitations you should know about before investing significant time. The companion AI cannot currently coordinate combined mantra combos with you in any reliable way. If you are hoping to set up chain triggers or timed skill rotations with your AI partner, you will be disappointed. The behavior tree does not support conditional chaining at this level. It reacts to your actions with a delay of roughly 1.5 to 2 seconds, which is functional for basic aggro management but completely inadequate for coordinated offense. Some players try to work around this by using very slow wind-up mantras and timing their own attacks to match, but this reduces both your damage output and your defensive responsiveness. The second major limitation is that companion behavior does not persist across different difficulty modifiers in a meaningful way. A companion trained on normal mode will not automatically adapt to hard mode enemy patterns because the behavior weights are tied to encounter difficulty parameters, not just raw combat frequency. I spent roughly 12 hours training a companion for hard mode only to realize the aggression weights were calibrated for normal difficulty enemy health pools. Retraining from scratch took another 20 matches to correct. If you plan to play on multiple difficulties, train on the highest difficulty you intend to use from the start.
Get the Full Details
For people who want a faster route, the only reliable shortcut I found is the mantra rotation method. Instead of locking into a single build, cycle through three different mantra combinations across consecutive matches while keeping track of which combination produces the behavior you want. After about 10 matches per combination, compare the results. The build that produced the cleanest decision tree is usually the one you should commit to for the long term. This approach cut my total training time from roughly 60 matches down to about 35 matches per companion. If the companion system is not working for you after an honest attempt, the alternative path is to focus on solo play with heavy mantra investment on your own character. Deepwoken is fully playable without an AI companion and several endgame encounters are actually easier solo because you do not have to manage positioning for another entity. The companion system is best viewed as a quality-of-life improvement for early to mid-game content rather than a mandatory feature for endgame progression.
