Everyone is writing about agent fleets. Almost nobody publishes the receipts, so here are mine for July.
Sixty-five work ledgers, forty-one of them actually finished — I checked each one against its artifact rather than trusting its own status field, and the labels were overstating by eight. Five hundred and sixteen commits in the ledger repo alone. One plugin published to npm on the twenty-first, three hundred and eighteen downloads in its first month. A public site that now re-exports and redeploys itself nightly, and only when the data actually changed.
Worked. Ledgers as ground truth, not session memory. Every worker is a headless session that reads a file, does the work, writes down what it did, and commits. When one dies mid-run — and they die — the next one resumes from the file rather than from scratch. The unglamorous part is what makes it survive.
Didn’t. I spent an evening on a database change that rewrote billions of rows, killed it, and then spent another hour cleaning up the dead tuples it left behind. Five separate merge plans before I understood that “the same vessel, inside this time window” is not a selective join key. The fleet was working the whole time. It was working on the wrong idea, quickly.
Next time. Make the agent prove the plan on a measured sample before it runs the plan. Speed without grounding is just a faster way to be wrong, and an agent that sounds certain is the easiest thing in the world to build.
That last part is the actual skill. Not prompting. Knowing when the confident answer is unearned.