Flue's agent is a function. The material names four hooks, not 16

The figure in play is 16 built-in TypeScript hooks. The supplied material does not contain it. What the withastro/flue README does contain is an example agent that uses four distinct hooks by name: useModel, useSandbox, useSkill and useTool ^[5][6]. No total is given anywhere in the text we have [c6a], so the honest version of this brief is four documented, count unverified.

The four are worth more than the missing sixteen anyway, because of how they are used. The example file is marked 'use agent' and exports a plain function, Triage, which calls useModel with a model string, useSandbox with a local sandbox, useSkill twice and useTool twice, then returns the task as a template string ^[2][7]. The README's framing is explicit: "The agent IS the function" ^[3], composing sessions, tools, skills, instructions, filesystem access and a sandbox to run in ^[1]. That is a harness assembled by executing code rather than declared as config, and the README does not say how often that code is evaluated [c6a].

That is the constraint. On DeepResearch Bench, across more than 500 agent sessions with 10,000-token system prompts, prompt caching cut API cost by 41-80% and improved time to first token by 13-31% across OpenAI, Anthropic and Google ^[8][9]. Losing the cache entirely therefore costs roughly 1.7x to 5x per call ^[12]. The same paper reports that naive full-context caching can paradoxically increase latency, and that what works is cache block discipline: dynamic content at the end of the system prompt, no dynamic traditional function calling, tool results excluded ^[10]. Benefits scale linearly across prompt sizes from 500 to 50,000 tokens and 3 to 50 tool calls, once the provider's caching minimum is met ^[11].

A framework where the prompt prefix is the return value of a function you rewrite is a framework that needs to guarantee that prefix is byte-stable between turns.