Why "just add an AI agent" keeps costing more than the plan on the whiteboard
As of July 2026, three different corners of the AI agent world surfaced the same complaint inside a 48-hour window. A Hacker News post titled "Who manages the agents?" pulled 72 points and 91 comments — that comment count alone tells you oversight isn't a hypothetical for the people running agents in production. A separate measurement post showed what an agent pays just to read: an average Wikipedia page costs 68,240 tokens of raw HTML, and Nike's homepage runs 353,000 tokens, before the agent has answered a single question. The same post noted that a built-in summarization step can cut a Wikipedia fetch down to roughly 950 tokens — a 98% reduction, if something strips the page before it reaches the agent's context. And a third builder, tired of hoping his agent would run its morning routine correctly, wrote a small declarative language, Skillscript, so a fixed sequence — check overnight tickets, summarize the deploy pipeline, flag anything urgent — runs the same way every day instead of depending on the model "getting it right" again.
Three different symptoms, one root cause: agents that improvise instead of running against rules someone wrote down. If you're deciding whether to build your own agent stack or hire a shop to run one, this is the checklist that decides whether the thing pays for itself or quietly eats your margin.
Why this matters
An agent with no fetch discipline, no memory policy, no written runbook, and no named owner will still work — that's the trap. It'll answer questions, close tickets, and draft emails, right up until the invoice or a bad decision makes someone ask "wait, who approved this?" By then you've usually paid for it twice: once in token spend, once in the cleanup. None of the four fixes below require a data science team. They require someone deciding, in writing, what the agent is and isn't allowed to do.
The 5-step playbook
- Audit what your agent actually fetches, not what you assume it fetches. Raw-HTML web research is the single biggest silent line item in most agent bills — a page that should cost a few hundred tokens can cost tens of thousands if nothing strips it down first. We wrote a full breakdown of where that spend goes in Your AI Agent's Web Research Bill Is 70x Bigger Than It Should Be; the short version matches the 68,240-vs-950 token gap above — a summarization step before content reaches your agent's context routinely cuts the bill by 95% or more.
- Write a memory policy before you pick a memory tool. Decide, in writing, what the agent is allowed to keep past a single session — a customer's stated preference, an open ticket number — versus what it must discard, like the raw text of a page it scraped an hour ago. Pick the storage mechanism (vector store, structured table, plain log) after the policy is written, not before. Choosing the database first is how teams end up storing everything and forgetting why.
- Turn the routine into a script, not a paragraph of instructions. A prompt is a suggestion the model reinterprets fresh every run; a declarative script is a rule the agent follows the same way each time. If your agent runs the same morning brief, ticket triage, or outreach sequence daily, write it down as explicit steps with defined inputs and outputs — the same discipline behind Skillscript's approach — so the fiftieth run behaves like the first.
- Cap tool and API calls per task. Set a hard ceiling — for example, no more than 5 web fetches or 3 database writes per task — before you turn an agent loose on anything customer-facing. A ceiling forces the agent to stop and ask for help instead of spending its way through an edge case it doesn't know how to handle.
- Name a human owner and put a weekly review on the calendar. "Who manages the agents?" generated 91 comments because most businesses running agents in production don't have a clean answer. Someone specific — not "the team" — should read a sample of what the agent did each week and sign off, the same way you'd review a new hire's first month on the job.
Common pitfalls
- Assuming a web-fetch tool is "free" because the API call itself is cheap — the token cost of what comes back is the real bill, and it can run 10x to 70x higher than the answer actually required.
- Letting memory grow without a discard rule, so the agent's context — and your storage cost — grows every week whether or not any of it is still useful.
- Treating the system prompt as documentation. If the agent's rules live only inside a prompt nobody has reread in months, you don't have a runbook, you have a guess.
- Skipping the ownership question until something goes wrong, instead of assigning it on day one alongside the budget.
Before you build it yourself, run this checklist
If you're weighing DIY against hiring a shop to build and run the stack, these five questions decide it fast:
- Do we know, in tokens or dollars, what each agent task costs today?
- Is there a written memory policy, or just "whatever the vector database defaults to"?
- Are recurring tasks scripted, or does someone re-prompt the agent from scratch each time?
- Is there a per-task ceiling on tool calls?
- Is there one named person who reviews agent output every week?
Two or more "no" answers isn't a reason to avoid agents — it's a reason to get the guardrails built by someone who's already hit these costs once and fixed them, instead of finding out what an unmetered agent costs on your own invoice.
Want a plain read on how your business actually shows up when AI tools get asked for the "best [category] near me"? Start with a free AI Visibility Report — it takes 24 hours and no guardrails are required to request it.