Why Do AI Agents Use So Many Tokens? How to Control Automation Costs

When Gavin Campbell and I discussed OpenClaw and Hermes on Home Gadget Geeks 680, we asked a question that was funny because it was a little uncomfortable: are AI-agent frameworks just elaborate ways to burn tokens? After months of running agents around my home lab and publishing workflow, I’ve learned that there’s a real concern behind the joke.

I’ve watched a simple task turn into an expensive chain of activity. One agent reads a report, another summarizes it, and a third searches for context the first two already had. The work might eventually get done, but that doesn’t make the process efficient. My biggest lesson is to stop asking which model is cheapest and start asking which parts of the work need a model at all.

Why Do Small AI-Agent Tasks Use So Many Tokens?

Tokens are the units of text a language model processes. The visible answer is only part of the work: requests can include system instructions, previous messages, tool descriptions, retrieved files, API responses and repeated attempts when something fails. A short answer can require a substantial amount of input context.

A check that runs every half hour might repeatedly load those instructions, call tools and interpret the same healthy state. Provider caching can help, but it doesn’t make repeated tool use or reasoning free. With a local model, the cost often shifts toward compute time, energy and resources rather than a per-token invoice.

My Jimbot Status Report Is a Good Place to Start

Every morning, Jimbot tells me whether my overnight backup succeeded, how much free space is left, whether Home Assistant is current and what deserves attention. Most of those facts don’t require creative reasoning. A backup finished or it didn’t. The version number matches the target or it doesn’t.

Those checks belong in predictable scripts. My backup process already has a schedule and retention policy. The script can count retained sets, check the latest run and report available space. AI becomes useful when there’s an exception to interpret, conflicting evidence to reconcile, or a decision to explain. I don’t need a fresh language-model conversation to rediscover the same healthy backup fourteen times.

A recent YouTube scheduling job gave me another example. An automated quota guard stopped because its local counter lacked a full-day baseline, even though the YouTube connections passed. The important distinction was between an exhausted quota and not enough evidence about the quota. I manually checked the API, authorized that one batch, and preserved the guardrail for future work. That’s the kind of decision process I want my agents to help with—not an excuse for them to keep querying the same thing.

Four Changes That Keep Agents From Repeating Work

1. Collect measurements once. One deterministic report should supply the facts that several managers need, instead of each agent rediscovering them.
2. Separate checking from explaining. File existence, version comparisons and time-based thresholds belong in code. AI adds value when context and judgment matter.
3. Keep task context focused. Supply the relevant log or report, not the whole project archive. Extra history can be expensive and distracting.
4. Match model capability to the task. A smaller or local model may be enough for classification. Hard debugging and editorial judgments may warrant a stronger model. Sometimes the best choice is no model at all.

These are operating principles, not a promised percentage of savings. Token pricing, caching and tool overhead differ across providers. What matters is being able to see which calls help and which calls merely repeat work.

Local AI Isn’t an Automatic Bargain

I’ve experimented with a Mac mini M4, an RTX 3060 desktop and hosted models. Local inference can offer privacy and more control over workloads, but it still costs power, time and machine capacity. A model with no per-request bill can still be a poor choice if it keeps hardware busy for a simple check.

I explored these trade-offs in my Mac mini M4 vs. RTX 3060 local AI comparison. My advice is to prove a useful workload first, then decide whether dedicated hardware actually helps.

Specialized Agents Can Save Effort—or Multiply It

Giving separate managers responsibility for content, infrastructure or other work has helped me define clearer boundaries. But multiple agents aren’t automatically more efficient. If each one independently polls the same source, reconstructs the same context and invokes a large model, more specialization can mean more overhead.

The better split is by responsibility: a script gathers evidence, an AI manager interprets ambiguity, and I approve consequential changes. I wrote more about those boundaries in Smaller Keys and Clearer AI Manager Lanes.

Measure the Result, Not Just the Token Count

I don’t want to optimize a system so aggressively that it becomes unreliable. A cheap report that misses a backup failure is not a win. Neither is an elaborate multi-agent discussion that never produces an outcome. I care whether the job finished, the evidence was trustworthy, and the system stopped when it should have.

Gavin Campbell and I explored the original token question in HGG680. There’s also a focused video about AI-agent frameworks and token usage. For your first experiment, my OpenClaw getting-started guide explains why limited permissions and one practical task beat a sprawling deployment.

The question I come back to is simple: Did that AI call produce something I couldn’t have gotten just as reliably from a script? When the answer is yes, I’ll gladly spend the tokens. When it’s no, that’s an opportunity to simplify the system—not build another agent. My practical test is to take one recurring agent job and note what data it retrieves, how often it calls a model, and what decision the model actually makes. Move routine checks into a script, then compare the result for a week. If you’re still choosing a framework, start with how I evaluate OpenClaw and Hermes for a home lab. For devices and sensors, see where I draw the line between AI advice and smart-home control.