Case file
A Python decorator that caps LLM API spend — marketed as "runtime authority for autonomous agents" — with four incident cards demonstrating the problem (including one where $1.40 in tokens caused $50K in damage while staying within budget), and a contact section where a founder reads every email, except bugs, which should go to GitHub.
Cycles adds budget caps and action gates to AI agent calls via a Python decorator: @cycles(estimate=5000, action_kind='llm.completion', action_name='openai:gpt-4o'). The product category: "runtime authority." The incident card for the $1.40 / $50K scenario notes the agent "stayed within its token budget the entire time" and advises "gate actions, not just spend." The contact section says a founder reads every email, then redirects bugs to GitHub.
Exhibit A — Evidence
Captured 2026-05-18
Score breakdown
Editorial roast
By Editorial Desk · Filed against runcycles.io↗
¶ 01
"Runtime authority for autonomous agents." The product: a Python decorator. `@cycles(estimate=5000, action_kind='llm.completion', action_name='openai:gpt-4o')`. This is a budget annotation on a function call. The marketing category for this mechanism: "runtime authority." A cap on API spend is authority. An action gate is enforcement. The blast radius of an email going to the wrong address is a governance problem. The vocabulary is technically defensible. The function is a decorator.
¶ 02
"200 emails. $1.40 in tokens. $50K in damage." The second incident card. The agent stayed within its token budget the entire time. The spending cap did not prevent the damage because the damage was not caused by overspending — it was caused by the action the agent was authorized to take. The card's advice: "Gate actions, not just spend." This is the product explaining, in its own marketing copy, the category of problem that its primary feature — the spending cap — does not solve.
“A Python decorator that caps LLM API spend — marketed as "runtime authority for autonomous agents" — with four incident cards demonstrating the problem (including one where $1.40 in tokens caused $50K in damage while staying within budget), and a contact section where a founder reads every email, except bugs, which should go to GitHub.”
¶ 03
"Auditor asks: prove the agent was under control." The fourth incident card assumes a regulatory auditor who asks for AI agent pre-execution authority trails. Standardized audit requirements for AI agent governance do not currently exist in most industries [redacted] or jurisdictions. The Cycles product is being sold as compliance infrastructure for a compliance requirement that is, as of 2025, approximately one or two legislative cycles ahead of where most enterprises are. The scenario is real. The auditor is optional.
¶ 04
"Most teams reach out for one of four reasons — they all land in the same inbox, and a founder reads every one." The contact section. Directly below: "Found a bug or have a concrete technical issue? Open a GitHub Issue instead — it's faster for everyone." The founder who reads every email is not the right person for bugs. The four reasons that land in the inbox are: piloting, failure modes, roadmap feedback, and evaluation. Bugs are a fifth reason, redirected. The founder reads every one except the ones that should go somewhere else.
— 30 —
More from the archive
An AI workflow automation tool whose hero defines itself as "automation that works" (implying competitors do not), whose value proposition is "intelligent automation — without the manual work" (the definition of automation), whose Zapier comparison claims to be "1k+ cheaper" without specifying the unit, and whose integration count of 6,000+ appears in at least two separate section headings on the same page.
A portfolio hosting product that says "30 seconds" twice in two consecutive section headings, positions its customers as "people, not developers" in a product named Supa-drop-.host, offers "beautiful templates" with beauty as the only descriptor, and closes with a footer asking if you are ready to use a product that, by its own repeated count, takes 30 seconds.
An anti-AI-slop writing tool — packaged in a hero / feature-cards / pull-quote / pricing-table / FAQ SaaS template that any AI would generate unprompted, decorated with a typewriter stock photo to signal authenticity, whose product logic is: AI writes too perfectly, so pay $9 a month to add the imperfections back.
Correction channel
Ask for a correction or update. Satire is more effective when the facts are not lazy.
