My personal opinion on this is it doesn't make sense to me that combat specifically between AI has to be subjected to the same level of physics collision reliance as the player. The player is the only one who appreciates the feedback from a heavily physics dependent combat resolution.
I feel like combat resolution between AI fighting could be made to use less physics/collision/hitbox reliance, and add a bit more CRPG style hit%, block% crit% auto-attack/ autoblock fighting when fighting amongst themselves, with values derived from troop stats, especially since physics is already particularly janky and makes agents shuffle and bounce around in mob situations, and collision being such a mess, creating for a very chaotic and inconsistent melee combat simulation. It will also remove the need for every single AI to constantly track its movement vectors vs target projected vectors, attack/block direction, attack animation timing calculations, and predictions. AI simply gets in range to target enemy AI, makes correct facing and distance adjustment, makes attack/defense rolls, resolved and reflected through a bit of movement and animation.
How much does it matter what goes on under the hood for resolution between AI fighting, when player won't see most of it in a large battle, since we're too busy fighting ourselves or giving orders while focusing more on troop formations and position anyway. I'd care more that my elite infantry is holding the line while fighting and defending effectively and my cavalry lancing enemy archers reliably, but it's hard to make the AI be reliable while subjecting it to a very unreliable mob physics, messing up its calculations with a slight bump, bad collision or a minor shuffle.
Perhaps the physics collision-based combat resolution could only apply to player (when striking at an AI agent) and AI targeting the player, where the impact and feedback of it matters.
