We introduce a measure called width, quantifying the amount of nondeterminism in automata. Width generalises the notion of good-for-games (GFG) automata, that correspond to NFAs of width 1, and where an accepting run can be built on-the-fly on any accepted input. We describe an incremental determinisation construction on NFAs, which can be more efficient than the full powerset determinisation, depending on the width of the input NFA. This construction can be generalised to infinite words, and is particularly well-suited to coBüchi automata in this context. For coBüchi automata, this procedure can be used to compute either a deterministic automaton or a GFG one, and it is algorithmically more efficient in this last case. We show this fact by proving that checking whether a coBüchi automaton is determinisable by pruning is NP-complete. On finite or infinite words, we show that computing the width of an automaton is PSPACE-hard. 1 Introduction Determinisation of non-deterministic automata (NFAs) is one of the cornerstone problems of automata theory, with countless applications in verification. There is a very active field of research for optimizing or approximating determinisation, or circumventing it in contexts like inclusion of NFA or Church Synthesis. Indeed, determinisation is a costly operation, as the state space blow-up is in O(2 n) on finite words, O(3 n) for coBüchi automata [16], and 2 O(n log(n)) for Büchi automata [17]. If A and B are NFAs, the classical way of checking the inclusion L(A) ⊆ L(B) is to determinise B, complement it, and test emptiness of L(A) ∩ L(B). To circumvent a full determinisation, the recent algorithm from [3] proved to be very efficient, as it is likely to explore only a part of the powerset construction. Other approaches use simulation games to approximate inclusion at a cheaper cost, see for instance [8]. Another approach consists in replacing determinism by a weaker constraint that suffices in some particular context. In this spirit, Good-for-Games automata (GFG for short) were introduced in [9], as a way to solve the Church synthesis problem. This problem asks, given a specification L, typically given by an LTL formula, over an alphabet of inputs and outputs, whether there is a reactive system (transducer) whose behaviour is included in L. The classical solution computes a deterministic automaton for L, and solves a game defined on this automaton. It turns out that replacing determinism by the weaker constraint of being GFG is sufficient in this context. Intuitively, GFG automata are non-deterministic * This work was supported by the grant PALSE Impulsion.