Cost and Runaway Loops

An autonomous agent works by looping: try, check, try again. That loop is the whole point — and it's also what can quietly run up a bill or spin forever.

This connects to the cost-safety idea from earlier: a flat-rate subscription tool fails by stopping; a metered paid-API agent fails by charging. When a coding assistant on a subscription hits its limit, it just won't do more until the window resets. When a script you wrote against a paid API gets stuck in a retry loop, it keeps calling — and every call costs money — until you notice or your card maxes out.

So the rules are blunt:

• Never run an unthrottled agent loop against a paid API. If you've written your own loop that calls an LLM API, it must have a hard cap — a maximum number of iterations, a spending limit, a timeout. No exceptions, no "I'll watch it." You won't.
• Give every loop an exit. "Stop after N attempts," "stop if the tests still fail twice," "stop if you've been running ten minutes." An agent without a stop condition is a bug, not a feature.
• Watch the first runs. Before you trust an agent to run in the background, watch a few foreground runs end to end. You're checking that it converges — that it gets closer to done, not stuck circling the same fix.
• Prefer the tool with the guardrails. A managed assistant with built-in limits is safer to leave running than a raw API loop you wrote at midnight. The subscription failing closed is a feature; the API failing open is a hazard.