I am afraid that @ucfagls is correct here, since as @ecologics pointed out and I have since been able to confirm in my own analyses, the brms implementation of GAMs does not support the automatic knot count selection as implemented in mgcv (which uses generalized cross validation for this purpose). To get around this, you could either manually reduce the number of knots (e.g., s(x1, k = 5)) if you believe this is appropriate, or you could follow @ecologics suggestion to use mgcv to estimate the ‘correct’ number of knots (see his reply to my previous question for details/caveats).
If you make any progress on this issue I’d be most interested in what worked - I am currently fitting approximately (20 x ) 100 GAMs on about 5k data points each, largely because the larger models are computationally intractable (at least for me at my current level of expertise). Even then fitting the 100 models takes 4 days on 4 physical (8 virtual) cores…