This paper identifies and fixes a surprisingly overlooked inefficiency in how LLM serving infrastructure handles agentic workloads. Most LLM serving systems cache the model's internal KV state so they don't have to recompute attention over a prompt they've already seen, but the authors point out that the tokenization step — turning raw text into token IDs — usually isn't cached the same way, so the full request text gets re-tokenized from scratch on every single call. That's a minor cost for a one-shot chat message, but coding agents don't work that way: they resubmit an ever-growing transcript after every small tool result (a file read, a test run, a grep output), so the same several-thousand-token prefix gets re-tokenized over and over as the conversation grows. Analyzing 153,951 real calls from two agent ecosystems, the authors found the typical call only appends about 1,400 characters of new content, yet at a 94.1% prompt-cache hit rate for the underlying model, tokenization alone can eat up to 64% of the time-to-first-token — meaning the bottleneck for a "cached" request has quietly shifted from model computation to a text-processing step nobody was optimizing. Their fix, TokTier, is a stateful tokenization service that maintains token-boundary state across calls, so it can incrementally tokenize just the appended text instead of redoing the whole transcript, while guaranteeing the output token IDs are identical to what full re-tokenization would have produced (avoiding subtle correctness bugs from mismatched boundaries near an append point). For developers running or building agent infrastructure at any real scale — not just frontier labs — this is a concrete reminder that as agent loops mature, the next round of performance wins moves out of the model itself and into the serving stack around it, and tokenization is a specific, fixable place to look for latency that's currently being wasted for free.