i personally haven't found that much bugs, simply because many things cannot be configured (as they are non-existant) and can't generate errors because of that
but there were many times of session-disconnects, no communication after the hand-shake, no propagation of routes. most times a restart fixed everything, but in a production environment you cannot restart a core-router once a week just because it stopped working
i tried to debug this a few times with mt-support, and many others did the same as you can read in quite a few threads on this forum, but mostly the only solution was a restart or an update which created other problems not only in the bgp-stack.
so for me this experiment is done, i won't try it until there have been some really really really promising case studies with traffic more than 300mbit/s and more than two peers. as openbgpd is a fine and stable software-router and juniper makes really very cheap routers there is no need to use mt desperately in every part of our network.
don't get me wrong, i like the software and most of the routerboards are fine, but there are parts where mt is good at (wireless, pppoe) and there are parts where mt (in my opinion) should have left the market for another company, as a bad protocol-implementation is worse than no implementation, especially in a thing like bgp, which the big mass of their customers is not needing, but is a backbone technology, so they cannot depend on their customers to beta-test it as they do with other technologies.
i can test a stupid-l3-switch in any part of my network without any problems, but a bgp router can only be tested in a realistic environment with working peers and live traffic, so i need to replace a maybe working bgp-setup to test a new product. that will happen only rarely i think