This Kills Your AI Token Usage

We need to talk about the biggest killer of your AI token allowance. It’s not necessarily the models you use, nor is it the different modes of thinking. It’s nothing to do with fast mode and is not related to context build up or dealing with complex tasks. It’s the loops!

Understanding the Issue

Agentic AIGlossaryAgentic AIAI tooling used as an active coding partner that can plan, edit, debug, and iterate across a project rather than only answering isolated prompts. has a certain level of autonomy. It’s one of the most powerful elements of using them for troubleshooting and feature iteration. It’s a great way to multitask, but it’s unfortunately the root of this token eating problem. Agents have gotten smart enough to know that they can monitor terminals and calls by polling.

It’s a great capability that allows the user to not get involved when it comes to external tooling. But it comes with great overhead. Every time the agent polls, it’s spending tokens. This compounds when you realise that the agent decides to poll almost every second to try and speed up the output time. This is ultimately what kills token usageGlossaryToken UsageThe amount of model context and generation capacity consumed while an AI assistant reads files, reasons, writes output, and revises work. limits and exhausts your spenditure before you get to do anything.

Fixing the Issue

It can be a little difficult when you’re used to agents running endlessly. It’s annoying that system prompts don’t already do this, but that’s why we have instruction files, pre-prompt instructions and other tooling to help create guard rails in place. This is the guard rail that I’ve added to my AGENTS.md global file to tackle this problem. I still need to be vigilante as agents don’t always listen to the instructions.

Do not poll terminals, loops, or background processes waiting for an action to occur.
Do not spend tokens on repeated status checks, sleeps, retries, or idle waiting.
If progress depends on an external event, user action, interactive prompt, or long-running process, stop immediately.
Then:
1. explain what is blocking,
2. give the user the exact command or script to run,
3. ask them to report back the result.
Prefer writing a script, test, or command the user can execute over monitoring the process yourself.

Retrospective

Knowing how capable and costly agentic AIGlossaryAgentic AIAI tooling used as an active coding partner that can plan, edit, debug, and iterate across a project rather than only answering isolated prompts. can be. It’s worth realising that the most effective use of tokens is to create tokenless automation from them. Instead of getting an agent to run a series of functions or tasks that is repeatably dumb, we can get them to generate scripts which do that for us. We can then either run these scripts manually, or create an Agentic Skill which gets the agent to fire and forget the scripts for us. Special note on the fire and forget element! We don’t want the agent to start babysitting as that would cost unnecessary token usageGlossaryToken UsageThe amount of model context and generation capacity consumed while an AI assistant reads files, reasons, writes output, and revises work..