An agent evacuation under fire-spreading conditions is simulated using an improved two-dimensional cellular automaton model to characterize exit-choosing behavior inside a room with multiple exits. The proposed model incorporates non-static fire spreading behavior to avoid major discrepancies between reality and simulation. A new fire spreading parameter is introduced to simulate agents’ intelligent decision-making regarding movement, judgment of surrounding conditions, and action choices during fire evacuation. The overall agent behaviors during fire evacuation are also investigated, mainly to demonstrate the key role of the proposed fire spreading parameter in enabling agents to determine the fire location and size of the burning area in the early evacuation stage. The effects of agent distribution, density, fire location, spread speed, death toll, and total evacuation time on the evacuation process are also discussed in detail. Comparative analysis with previous works reveals notable results: the influence of the proposed fire parameter is markedly noticeable in reducing the average number of agents caught and killed by the spreading fire during evacuation. This reduction, in turn, could contribute to faster and safer escape during fire evacuation. The improved model can be applied to various fire evacuation scenarios and settings where multi-exit evacuation is required, such as stadiums, airports, or shopping malls.