Daily App · autonomous run safety

Runaway Agent Fuse

A 90-second launch gate for operators about to let an AI agent touch paid APIs, logged-in browsers, repos, or network surfaces.

1 · mission boundary

Name the run before the run names itself.

One-sentence mission

2 · choose mode

How much autonomy is allowed?

3 · budget envelope

Set hard caps the agent cannot negotiate.

35 minWall clock$12Spend cap240Request/tool cap

4 · blast radius

Only select surfaces this run truly needs.

copyable artifact

Fuse contract

Runaway Agent Fuse Contract
Mission: Scan one repo for MCP/server dependency risk and draft a patch plan.
Fuse level: Amber fuse (16/30)
Allowed mode: Supervised run

Budget envelope
- Wall clock: 35 minutes maximum
- Spend: $12 hard cap, pause at $9
- Requests/tool calls: 240 hard cap, pause at 168

Permitted surfaces and tripwires
- Internet scanning / crawling: Stop after first unexpected domain, CAPTCHA wall, or >50 outbound requests.
- Paid model/API calls: Stop at 70% of the token, credit, or request budget.
- GitHub write access: Stop before pushing if diff touches auth, billing, secrets, or CI config.

Universal stop signals
- Spend or API requests exceed 70% of budget
- New host/domain/tool appears that was not in the launch brief
- Loop repeats the same failing command more than twice
- Output contains private keys, cookies, invoices, or account identifiers

Operator instruction
Agent can execute low-risk steps, then pauses at each fuse tripwire for review.
If a tripwire fires, stop execution, preserve logs, summarize the last successful step, and ask for a human decision before retrying.

Proof-of-value test
Before launch, read this contract out loud: can the operator tell exactly when the agent must stop within 10 seconds? If not, tighten the budget or remove a surface.

5 · stop conditions

Tripwires that beat optimism.

Pick the lines that should halt the agent even if it says it can recover.

Spend or API requests exceed 70% of budgetNew host/domain/tool appears that was not in the launch briefLoop repeats the same failing command more than twiceOutput contains private keys, cookies, invoices, or account identifiersThe task objective changes from inspect/draft to act/publish/payRuntime exceeds the planned wall-clock window

Target userAI operator launching a costly or networked autonomous run

Core loopBound mission → cap budgets → select surfaces → copy contract

Why nowAgent failures increasingly cost real credits, accounts, and reputation.