追踪 agent 44 次调用：84% 是决策而非写作

kiruwaaaa @kiruwaaaaaa原推发布于 2026.10.13

在 X 上打开原推

holy sh*t, i traced 44 calls from one agent run and 37 of them never wrote a word illustrative trace of a support-triage run, every call tagged with one label > write: it ends up as text a person reads > decide: it only picks the next branch, then gets thrown away here's how the 44 split: > 7 writes: reply drafts, a plan, a summary > 37 decisions: which step next, is this source useful, is the draft ready, which tool, escalate or not that's 84% decisions, and it's not unusual for agent setups what actually got me is how most of us run it. we pay writing prices for decision work, a prose model spends paragraphs on a fork that already had a known answer the fix is a split > claude creates > jev picks, scores and gates > code does the thing a decision comes back as a typed answer with a probability in 70 to 500 ms (vendor figures), so your code acts on high confidence, reviews the middle and escalates the low try it tonight: log one run of your agent, tag every call write or decide, count save this for your next agent build how many of your agent's calls actually write something? full 24/7 stack in the article below ↓

这条是引用着另一条写的：x.com/kiruwaaaaaa/status/2109593981122195661