Retries inflated one agent turn to 47,004 input tokens for a 280-token answer

A tool call inside a hosted skill is not a cheap local function call. Each one is a full model round trip that carries the system prompt, the tool schemas, the chat history, and every tool result that came before it ^[7]. Call four tools in sequence and the fourth call pays for the first three results plus everything already sitting in the thread ^[3][7]. Steef-Jan Wiggers, who wrote the post ^[21], put the diagnosis plainly: "Not one line of my tool code was wrong," he wrote ^[10].

One turn spent about 168 input tokens for every token it produced ^[17]. At that input size, a 500K TPM deployment covers roughly ten turns a minute for the whole application: 500,000 divided by 47,004 is 10.6 ^[18]. Raising capacity bought a different error message ^[5].

The evidence the agent needs is cheap and deterministic. RDAP returns a domain's registration date, registrar and age; a direct HTTPS fetch confirms the site is reachable and checks whether it links contact, about, terms, privacy and returns pages; the Internet Archive CDX API shows when the site first appeared and whether its history is continuous ^[12]. The RDAP helper is an httpx GET against rdap.org with a 15 second timeout that returns registered false on a 404 ^[16]. "Nothing clever, which is the point," Wiggers wrote ^[19].

Reputation was the part that could not be gathered without a key. The Trustpilot content API needs a paid business account and scraping review sites violates their terms, so the second tier runs Tavily searches against site:trustpilot.com, the business name plus "reviews", and the business name plus "scam OR fraud OR oplichting" for Dutch shops ^[15][13]. The model sees result titles, URLs and snippets, and never invents a rating ^[13].

When no key is configured, that tier reports itself unavailable and the agent says so in its verdict; graceful degradation was a design goal ^[14]. That choice is what keeps a missing key from becoming a quota incident. An error return invites another attempt; an unavailable field is data the model can report and move past ^[9][14].

For the 47,004 to transfer to your project, two things about this setup have to hold on yours. The runtime keeps the thread for you, so failed turns and retries stay in it ^[8]. And at least one tool has to fail in a way the model reads as worth retrying ^[9]. The redesign the post lands on is one sentence: a single deterministic tool gathers every signal, under the pattern Wiggers calls "evidence in code, judgment in the model" ^[11]. The account stops there, with no agent_token_usage line for the rewritten version ^[20].