When I wrote infer_tree for phylokit-mcp, the first version had a bootstrap flag that defaulted to off. Bootstrapping is slow, and I figured an agent asking for a quick tree would not want to wait for it.

Then I tried it on a bad alignment. I simulated sequences from a 7-taxon tree I knew, cut the alignment down to 60 sites, and ran the tool. The tree came back fully resolved and looked fine. One of its clades is not in the tree the data were simulated from.

The Newick string gives no hint of that. The only thing in the output that flags the problem is the bootstrap value on that clade, and with the flag off there is no bootstrap value. An agent reading the result cannot tell a solid tree from a bad one, and the person reading the agent’s summary is one step further removed.

So I removed the flag. infer_tree always bootstraps and always returns per-clade support.

The other servers follow the same idea. plantcv-mcp returns the segmentation mask with every set of trait values, because a bad mask still produces a plausible leaf area. breedsim-mcp runs replicates and reports the spread; five seeds of the same three-cycle programme gave genetic gains with a standard deviation of 0.247, so a single run quoted to three decimals is mostly noise.

Anyone doing these analyses by hand already knows this. The difference with an agent in the loop is that the person asking may never look at the intermediate output, so the tool has to attach the evidence itself.