The work log for that morning says I ran a fleet of about 30 concurrent runs “after the author noted main-context iteration wasn’t burning quota fast enough.”
Thirty runs at once
The previous night’s session took the blog queue from 2,635 down to 38 and produced 106 AI-drafted posts, 16 newsletters and 19 social posts. They went through multiple codex review rounds and a framework audit, and the first 8 went live. The log says main-context iteration wasn’t burning quota fast enough.
So I fanned out to about 30 concurrent runs. By the end of that session the live set had gone from 8 to 63 posts, and every new one had twitter and linkedin variants. The session ran 7h 26m.
That was the stretch where I lost track of what I was spending.
The accounting
The weekly retro puts it in one sentence: parallel Opus sessions burned quota faster than I tracked during the blog push.
My work log does record tokens and wall time for single calls, like codex exec (medium, 95614t, 364.6s). A serial loop gives you a gauge you can read between iterations. A fleet removes the gaps between iterations, so the spend shows up after the fact.
The log also records research on a workstation purchase because the 11-year-old laptop was choking at 10-15 agents. The fleet was about 30. That note puts the limit at 10-15 agents.
Volume against a number I could count
The same morning’s log shows four codex framework findings closed: a .png.png OG image regression, an unmounted CodeCopyButton, hardcoded SystemPage colors, and broken tag URL encoding.
The week came to 1,260 commits across 23 repos, 174 sessions, and 533 commits on Wednesday alone. A big number and a controlled number are different things. The commit count is exact. The retro says the quota burned faster than it was tracked.
The workstation is researched and priced but not bought, so the laptop is still the ceiling.
A fleet sets your rate of spend. Mine was about 30 concurrent runs, and the quota burned faster than it was tracked.
AI Skills
Use this lesson with the AI assistant you already use
After a serial night that cut the blog queue from 2,635 to 38, I fanned out to about 30 concurrent runs because single-context work wasn't burning quota fast enough. The live set grew from 8 to 63 posts over a 7h 26m session, but the quota cost went untracked.
Paste the prompt, share only the context needed to answer it, and treat the result as a draft for your review. Do not include confidential information or let an AI assistant make changes without your approval.
Optional: for a visual report and saved memory, run /dxdev first.
Don’t have it? Get it at dxdev.com/skills/dxdev. The prompt works without it.
dxdev LESSON · paste into your AI coding agent
LESSON: A Parallel Fleet Sets Your Rate Of Spend, So Cap It And Meter It Before You Launch
SOURCE: dxdev.com/blog/2026-05-28_quota-arithmetic-of-parallelism
WHAT HAPPENED: I produced 106 drafts, 16 newsletters and 19 social posts in one main context, then decided that pace was too slow. I launched about 30 concurrent runs, and the live set went from 8 to 63 posts. My work log records tokens and wall time per single call, but a fleet removes the gaps between iterations where I would have read that gauge. I had also researched a workstation because the 11-year-old laptop choked at 10-15 agents, and I ran about 30 anyway. The week ended at 1,260 commits across 23 repos, and I could state that count exactly but could not state the quota cost. The retro admits parallel Opus sessions burned quota faster than I tracked.
THE RULE: Concurrency sets your rate of spend, so decide the spend limit and the concurrency cap before launching a fleet, not after. Track the cost in the same units as the budget, because output counts like commits do not tell you what you spent.
CHECK MY CODE, then report PASS or FAIL with file:line for each:
1. Before launching parallel runs, is there a written concurrency cap that is at or below the known hardware ceiling (here, 10-15 agents on the laptop)?
2. Is there a quota or token budget for the fleet run, with a spend figure that can be read while it is running and not only after it ends?
3. Does the end-of-run report include cost in budget units (tokens or quota) next to output counts such as posts or commits?
THEN PRINT: a table (check, PASS/FAIL, evidence, fix) + a verdict (applies / partially / OUT_OF_SCOPE / no) + the single most important next action.