Context steering
A movement AI that weighs every candidate direction against an interest-and-danger map, instead of summing separate steering forces the way classic steering behaviours do.
Classic steering behaviours add up vectors for "seek", "avoid" and "separate"; the trouble is that opposing forces can cancel and leave an agent frozen between two obstacles. Context steering instead samples a fan of directions around the agent and fills two maps — an interest map (how well a direction points at the goal) and a danger map (how close an obstacle) — then picks whichever direction looks best.
It threads tight gaps and streams of units more gracefully than the naïve sum, without a per-goal flow field to compute. Andrew Fray described it in 2013 and it spread through Game AI Pro 2 (2015). It is cheap enough to run per frame for hundreds of boid-like agents.