Sketch drawing recognition is knowledge-intensive since the system must know what to look for in a drawing. In earlier work, we identified 24 different types of representations, termed graphic units. For recognition of graphic units we combine a multi-agent approach and online recognition. Each agent is specialised for one graphic unit, continuously parsing the online input stream for typical stroke features. When a threshold is reached, the agent puts a claim. Each agent's cluster of features can be viewed as distributed over a decision tree. The activation pattern of feature clusters indicates which graphic unit is identified. We present the exhaustive set of features for agents and a binary decision tree over which the features are distributed.