As of July 2026, if your agent stack "just Googles it" every time it needs a fact, you are paying for that laziness in tokens, not intelligence. A developer who measured what Claude Code's built-in web fetch actually costs found that a typical Wikipedia article runs 68,240 tokens of raw HTML before any summarization happens. Nike's homepage: 353,000 tokens. Claude Code's own fetch tool gets the Wikipedia page down to roughly 950 tokens through summarization — a 98.6% cut — but the moment a site blocks the easy path (Cloudflare, a login wall, a JavaScript-rendered page), the agent falls back to the raw dump and the bill spikes right back up.
That gap between "950 tokens" and "353,000 tokens" is the whole story for any business running its own agent build. If you're standing up an agent to watch competitor pricing, pull reviews, or check what ChatGPT and Perplexity say when someone asks for the best plumber or dentist in your city, the fetch-and-read step is where your token budget actually goes — not the reasoning step everyone designs for.
The fix isn't a bigger model — it's less improvisation
The same week, a developer named Scott shipped Skillscript, a small declarative language for telling a local agent exactly what to do — check overnight tickets, summarize the deploy pipeline, flag anything urgent — the same way, every day, instead of hoping the model reinvents the routine each morning. It's a niche tool, 11 points on Hacker News, built to scratch one developer's itch. But the itch is the same one every business owner running an agent hits eventually: an agent that "figures it out" fresh each run is both expensive and unreliable, because every improvised decision is a token spend and a chance to drift from what worked yesterday.
Put the two stories together and you get a clear rule for anyone building or buying an agent stack in 2026: routine, repeatable steps — fetch this page, check this feed, format this report — belong in a script the agent runs, not in a prompt the agent re-derives. Reasoning is expensive and should be reserved for the parts of the job that actually require judgment: deciding what a competitor's price change means, drafting the outreach message, choosing which lead to call first. Fetching and parsing a web page is not that kind of task, and treating it as one is how a "simple" daily agent ends up costing more than the human process it replaced.
What this looks like in a local business's agent stack
Say you're running an agent that checks, daily, what AI answers show up when someone asks "best HVAC company near me" across ChatGPT, Claude, and Perplexity, then cross-references it against three competitor sites. Done the naive way — full page fetch, full context window, re-prompted from scratch each run — that's easily 300,000+ tokens a day before the agent has produced a single insight. Done with a caching layer, summarized fetches, and a deterministic script for the repeatable steps, the same job runs on a fraction of that, and the output is consistent instead of subtly different every morning depending on how the model happened to parse the page that day.
We've written before about how token and memory costs quietly become the biggest line item in an agent's bill, and this is the same problem wearing a different hat: the cost isn't the model, it's the volume of raw, unprocessed material you're feeding it because nobody engineered a cheaper path first.
The build-vs-hire math
If you're weighing building this in-house against hiring it out, the honest answer is that the model API call was never the expensive part. The expensive part is the engineering underneath it — the caching, the summarization step, the deterministic scripts for the boring 80% of the job — and that engineering is invisible until your first real bill arrives. A DIY agent that skips it will work fine in a demo and then quietly cost 50-100x more per useful answer once it's running daily against real websites.
What this means if you're weighing AI marketing or an agent build: check whether your agent (or your vendor's) is fetching raw pages every time or working from a cached, summarized layer — that one design choice is the difference between a sustainable daily agent and one you'll shut off in three months over the bill.
Want to know what AI is actually saying about your business right now, without burning your own token budget finding out? Get a free 24-hour free AI Visibility Report.