Human Control

Cost Management

A loop spends tokens the way a leaky tap spends water: quietly, on a schedule, and without end — unless somebody owns the meter. That somebody is you.

A loop spends tokens the way a leaky tap spends water: quietly, on a schedule, and without end — unless somebody owns the meter. That somebody is you.

The hook

Mid-afternoon on Day 1 of building this course, the page-writer loop pulled its own emergency brake: ≈79% of the daily token cap, 6 items left, switching to report-only. No human was staring at the meter — the meter was watching itself, because a budget rule had planted a tripwire right there. The human read the alert, raised the caps on purpose, and the loop wrapped up. That alert is still sitting in shared/loop-budget.md — cost management, caught in the act of working.

The practice (plain English)

Budget per loop, cap per day. Every loop declares its max runs/day and max tokens/day before it ever runs the first time — a loop with no cap is just a blank check with a heartbeat. Keep the numbers in one shared file the entire fleet reads.
The 80% tripwire. Cross 80% of any cap and the loop drops to report-only and announces it. (Measure against the cap as it reads right now — this repo once sailed to 90% because the tripwire was checking a stale number. The fix was a single line in the budget file.)
Know your loop economics. Read-only checkers run an order of magnitude cheaper than makers — this repo's Day 1 checker spent roughly 3% of its budget while the maker chewed through about 90% of its raised one. The corollary is handy: when in doubt, add a checker, not a fatter maker beat.
Estimate, then measure. Before a loop runs, do the back-of-envelope: rough tokens/beat × beats/day. After a week, the run log's tokens_estimate column tells you what it actually costs — and whether it's earning its bill.

The meters in each tool

Each tool hands you a per-session meter. Fleet-level accounting is the run log, summed per loop:

# Claude Code — live docs: https://docs.claude.com/en/docs/claude-code
/cost                     # current session spend
# Fleet accounting = your run log, by design:
#   every beat appends {"tokens_estimate": …} — grep and sum per loop.
# Budget discipline as a skill: this repo's loop-budget skill reads
# shared/loop-budget.md before and after every run.

# OpenCode — live docs: https://opencode.ai/docs
# Per-run usage appears in output/logs (see live docs); fleet-level is
# the same run-log discipline:
grep step-writer shared/loop-run-log.md \
  | grep -o '"tokens_estimate": [0-9]*' | awk '{s+=$2} END {print s}'

[!NOTE] Going deeper: cost is never only tokens — it's also your attention (the reports you have to read) and CI minutes (loops that get promoted into workflows). The same cap-and-tripwire shape covers all three. The fleet-level budget arithmetic — priority order, which loop to pause first — lives in multi-loop.

Check yourself

Q: Your nightly loop costs 60k tokens, and in three weeks its report has changed your behavior exactly zero times. The cap says it may keep running. Should it?

Answer

No — the cap only ever answers "may it spend?", never "is it worth it?" A report nobody acts on is a loop whose value is zero while its cost compounds night after night. Retire it, shrink its cadence, or change what it reports until it actually changes decisions. Budgets stop runaway cost; only the engineer stops pointless cost.

Try With AI

Work out this repo's real Day 2 bill: have your agent sum tokens_estimate per pattern from shared/loop-run-log.md, then hold it up against the caps in shared/loop-budget.md. Which loop came closest to its tripwire? Would you have set the caps differently — and what would you set for your first loop, given these actual numbers?

When it goes wrong

Symptom	Cause	Fix
Surprise bill at month's end	No caps, or caps nobody reads	Caps in one shared file, read at every beat start
Loop blew past its tripwire	Tripwire measured a stale cap	80% of the cap as currently written — re-read it each beat
Fleet starves the important loop	No priority order in the budget	Rank the loops; pause lowest-value first (labs before checkers)
Caps raised in a panic mid-run	The alert had no process behind it	Alert → human reads context → deliberate, logged raise (like Day 1's)

Glossary terms used on this page: token budget, tripwire, report-only, kill switch — see the glossary.

Sources: cost discipline draws on Panaversity's Loop Engineering: A Crash Course (S1) and Addy Osmani's Loop Engineering (S5). Full attribution: resources/sources.md.