The failure mode nobody warns you about is drift. Give an agent a clear skill — a format to follow, a checklist to run, a rule to never break — and it will honor it beautifully for the first few dozen items. Somewhere deep into a large dataset it quietly stops. Not dramatically, not with an error, just a slow slide back toward whatever it would have done without your instructions.
I have been measuring this rather than guessing at it: run the same skill across thousands of records and chart adherence against position in the run. The curve is the interesting artifact. Once you can see where the falloff starts, the fixes get obvious — re-anchoring the instruction, chunking the work, checkpointing state outside the context. This is the work I find most underrated in agent engineering right now.