Everyone argues about which model to use. Almost nobody talks about what happens between the moment an agent decides to look something up and the moment it actually reads the answer — and that gap is where a surprising amount of your AI budget quietly disappears.
A single web page can cost 68,000 tokens
A developer who uses Claude Code daily measured what it actually costs an agent to read a web page, and the numbers are worth sitting with. An average Wikipedia article, pulled as raw HTML and counted with tiktoken, runs about 68,240 tokens. Nike's homepage: 353,000 tokens. That's before the agent has done anything with the information — it's just the cost of looking.
Well-built tools compress that dramatically. Claude Code's built-in web-fetch tool summarizes that same Wikipedia page down to roughly 950 tokens, and it can clear Cloudflare bot-checks on sites like Indeed and Ticketmaster that would otherwise block the agent outright.
That gap — 68,000 tokens versus 950 — is the difference between an agent that is cheap to run and one that quietly burns through your API budget looking things up nobody asked it to look up so thoroughly. If you are paying per token, and with agentic workflows you are, the tool an agent uses to fetch a page matters just as much as the model sitting behind it. Most off-the-shelf agent builders never surface this choice to the business buying the agent, which is exactly why it goes unmanaged.
Memory has the same problem, it is just less visible
The same week, another piece made the case that memory strategy for AI agents deserves the same deliberate design as anything else in the system, not a default you bolt on and forget about. An agent that re-reads its entire conversation history every turn pays a token tax similar to fetching raw HTML. An agent with no memory at all forgets your customer's name three messages later. Neither is wrong exactly — they are just mismatched to the job.
The right setup depends entirely on what the agent has to do. A one-off research agent does not need to remember last Tuesday. A support agent handling repeat callers absolutely does. Treating memory as one-size-fits-all is how teams end up either paying for context nobody uses, or shipping an agent that feels amnesiac to the people relying on it.
The pattern: agents are only as efficient as their plumbing
Put these two stories side by side and a pattern shows up. The model gets the headlines, but the actual cost and reliability of an agent gets decided by the unglamorous infrastructure underneath it — what it fetches, how much of that it keeps, and how long it holds onto anything. Skip that design work and you get an agent that is technically smart but expensive and forgetful in practice. It is the AI equivalent of hiring someone brilliant who insists on reading every source document cover to cover, then forgets your name by lunch.
It is also why a weekend demo of “wire an LLM to a search tool” looks so different from an agent that survives contact with real customers and real invoices for months on end. The demo does not care that it burned 68,000 tokens confirming one fact it could have gotten in 950. Your monthly bill does, every single day the agent is running.
What to ask before you buy an agent
If you are evaluating an AI agent vendor, or wondering why your current one is more expensive than expected, these are the two questions worth asking before you sign anything:
- How does it fetch information? Raw page dumps versus a summarized fetch can be a 70x difference in token cost for the exact same lookup, and that cost compounds every time the agent runs.
- What does it actually remember, and why? “It has memory” is not an answer. The right memory strategy for a lead-qualification bot is not the right strategy for a support agent handling the same customer for a year. If nobody can explain the choice, nobody made one — it is just whatever the framework defaulted to.
What this means for AlphaForge clients
When we scope a build, fetch behavior and memory design are line items we decide on purpose, not defaults we inherit from a framework. That discipline is the difference between an agent that costs pennies per interaction and scales, and one that quietly eats your margin every time it looks something up.