<p>Many animal groups form structures such as flocks and swarms. However, how can the individual agents reconcile the simultaneous requirements of local collision avoidance, alignment, and group cohesion to achieve coherent collective motion? Here, we propose a minimal flocking model, where each agent is capable of vision-based steering interactions to achieve these (conflicting) goals. Numerical simulations in two dimensions show that local collision avoidance acts as a source for emergent noise and induces an order-disorder transition, triggered by the fast response of the flock to local directional changes. The emergence of large vortices at the critical point hints at a Berezinskii-Kosterlitz-Thouless-like transition. Deep in the ordered phase, the cohesion acts like a surface tension, favouring compact flock shapes. The competing interactions lead to pronounced shape and density fluctuations of the flock. These large fluctuations can be important for a fast response to external cues, which aids predator evasion and foraging.</p>