Real-time multi-agent collision-avoidance algorithms comprise a key enabling technology for the practical use of self-organising swarms of drones. This paper proposes a decentralised reciprocal collision-avoidance algorithm, which is based on stigmergy and scalable. The algorithm is computationally inexpensive, based on the gradient of the locally measured dynamic cumulative signal strength field which results from the signals emitted by the swarm. The signal strength acts as a repulsor on each drone, which then tends to steer away from the noisiest regions (cluttered environment), thus avoiding collisions. The magnitudes of these repulsive forces can be tuned to control the relative importance assigned to collision avoidance with respect to the other phenomena affecting the agent's dynamics. We carried out numerical experiments on a self-organising swarm of drones aimed at fighting wildfires autonomously. As expected, it has been found that the collision rate can be reduced either by decreasing the cruise speed of the agents and/or by increasing the sampling frequency of the global signal strength field. A convenient by-product of the proposed collision-avoidance algorithm is that it helps maintain diversity in the swarm, thus enhancing exploration.