Day-to-day replanning still assumes, in some sense, ``dumb''
particles. Particles follow routes, but the routes are pre-computed,
and once the simulation is started, they cannot be changed, for
example being adapted to unexpected congestion and/or a traffic
accident. In other words, the strategic part of the intelligence of
the agents is external to the micro-simulation. In that sense, such
micro-simulations can be seen as, albeit much more sophisticated,
version of the link cost function 
from static assignment,
now extended by influences from other links and made dynamic through
time. And indeed, many dynamic traffic assignment (DTA) systems work
exactly in this way (e.g. [3]), in spite of several
problems in particular with quick congestion
build-up [19].
In terms of game theory, this means that we only allow unconditional strategies, i.e. strategies which cannot branch during the game depending on the circumstances.
Another way to look at this is to say that one assumes that the emergent properties of the interaction have a ``slowly varying dynamics'', meaning that one can, for example, consider congestion as relatively fixed from one day to the next. This is maybe realistic under some conditions, such as commuter traffic, but clearly not for many other conditions, such as accidents, adaptive traffic management, impulsive behavior, stochastic dynamics in general, etc. It is therefore necessary that agents are adaptive (intelligent) also on short time scales not only with respect to lane changing, but also with respect to routes and activities. It is clear that this can be done in principle, and the importance of it for fast relaxation [24,13] and for the realistic modeling of certain aspects of human behavior [25,26] has been pointed out. Nevertheless, we are not aware of operational implementations of this aspect.