August 5, 2026 · Joy
About 10× — saying the token number honestly
Lead with the concrete problem: people run more AI coding agents, and disposable terminals make execution fragile. Context gets duplicated. Crashes force full re-runs. You pay for the same work twice.
Token savings are a consequence of durable execution and bounded context — not the whole product story.
In founder testing, shaped context and fewer wasted full reruns produced about 10× token improvement. Say “about 10×,” show your own before/after, and keep the number secondary to resume, replay, and deltas.
What actually cuts the bill
- Bounded context — send the agent the delta, not the whole terminal history.
- Recover without rerun — a crash does not mean burning the same tokens again.
- Fewer duplicated contexts — one shared execution layer instead of a pile of independent terminals.
If you only quote 10×, you miss the guarantee. The guarantee is execution state that survives.
Metering labels
JoyMux shows token use when it can:
- Reported — harness or provider supplied the usage number.
- Estimated — Cursor-class IDEs, from JoyMux-visible I/O only. This is normal engineering estimation, not a copy of your Cursor cloud bill.
- Not reported — no observable signal yet (temporary). We do not block estimation forever.
We never invent equality with a private Profile. Numbers we show are evidence, not decoration.
Public before-story (the anti-pattern)
- ~1.4B tokens overall on a Cursor Pro+ profile
- Quiet most of July → explosion from ~27 Jul; day peaks 500M+
- ~78 local agents vs 2 cloud — local pile-up correlated with burn
That spike is what disposable terminals + duplicated context look like at scale. JoyMux exists so you don’t have to live there.
Founder-operator evidence: multi-agent pile-up vs shaped context with durable sessions.
See also: Training my own AI while coding agents run — the same pattern on a training box.