Here are some more thoughts on automated strategic droids. In a science fiction context, this can make stuff seem more "real". However, in the real world, it will probably be quite a while until we entrust droids with strategy and combat to this extent.
We've previously discussed how independent or centralised a droid should be under certain circumstances (or have we? I don't remember, and I'm far too lazy to look it up). However, it seems likely that a combination of the two should be considered.
I feel as though a decent number of computer scientists would try to go for a 'cloud', where each droid would both calculate its own choices and contribute to a consensus result and go from there. It certainly sounds rather organic (as it were). However, centralisation and decentralisation both have their advantages, and should be used appropriately. There is also the weirder case when the droids are commanded to act without communication for a while, meaning they are under central control but cannot communicate with said control until some parameter is met (for stealth reasons). While an AI might come up with a solution, we can lay out some basic rules here for when to use either rule.
There are some cases where using decentralised control is more useful. Where time is of the essence, and communication takes much more time than processing, then a unit should be decentralised. In this clear case, we will call what the unit does a "reaction". The size of a unit could be a single droid, or a platoon. In these clear cases, a reaction should only be considered if communication time severely limits the success chance of that unit (or communication is unreliable).
There are some cases where using centralised control is more useful. Broad data processing for multiple units (say, ten thousand) would be much better done by sending the data to a few central units which themselves would be appropriately far enough away from combat that they don't have to react to things (thus being able to use that processing power for command). Assigning objectives (especially common ones) would also be better done further up the command chain rather than from the perspective of any single bot (or a cloud of).
Programs are relatively small, so each droid could have a potential "commander program", waiting in the wings for when the army is decapitated. The German Army found that when they trained all of their soldiers to be able to assume the rank immediately above them, the army itself was a lot more resilient to the kinds of shocks that an army takes over a war. Thus, you could easily choose a new commander.
You'd probably want to choose a commander (or set of commanders) far away from the front lines, but still has stable connections to a large number of troops.
The commander's role would probably be identifying likely routes for enemy supply based on current enemy positions (both likely and known) and terrain, and then identifying the most likely way to attack those supply routes. Frontline units themselves should not be concerned with that, because in a tactical situation supply routes are not quite as important (not entirely unimportant, but not as definitively advantageous). Or identifying its own routes of supply, what would be the most likely point of attack based on some disposition of the enemy, and then protecting those.
The commander would also have to do all of the logistical management that often comes in armies. This sort of thing would be relatively cheap, as it is the sort of task computers are far better at than people.
The commander would also have to collate data from all the front line units to determine likely strengths of unknown enemy formations (often reserves). It would use a combination of equations, learning, and traditional military doctrines to attempt to determine this (specifically how is for smarter people than me).
These are all things which work much more efficiently with a centralised command structure, though there are ways of them being accomplished clouded. However, centralising makes the most use of the resources in those cases, and nothing precludes using less active computers for processing things in a cloud like manner for particular task.
Anyway, just some more thoughts.