Seven ways to spend fewer tokens when coding with AI agents

A fact that surprises many people is that before your coding agent writes a single line, it’s often already spent a pile of tokens trying to find its way around your project. It opens a file, reads it, backs out, then opens another, wandering the hallways before finding the right room. 

That wandering is where a lot of your token budget gets used up. Fortunately, while inconvenient, you do have a lot of control over it. Most of your token savings come from seven core habits, and none of them require you to add a new tool or change your setup. 

1. Point the agent straight to the file

If you know which file has the problem, tell your agent early on. "The login logic is in auth/login.js" beats "fix the login bug" by a mile. If you’re vague about your instructions, the agent has to track down the file first, and might open five wrong ones on the way. Handing it the address is kinder to your token count, and kinder to the agent too.

2. Keep a short project map

Add a small file (many people call it CLAUDE.md, or just use a good README) that explains where the important files are and how to run the tests. The agent reads it once and stops rediscovering the same facts every session. Think of it as the note you leave for a house sitter so they aren’t texting you about where the spare towels are.

3. Start a fresh chat when the task changes

Long conversations include everything from earlier, and you pay for that history on every new turn. When you move to a different task, open a clean thread. Keeping one giant conversation going is a bit like hauling every grocery bag from the whole month around the store with you.

Here’s what that looks like in real money. Leave a coding agent open and idle for a couple of hours and its prompt cache often expires. That cache is the stored copy of your context that keeps repeat requests cheap, so once it’s gone, your next message has to rebuild it from scratch. In the screenshot below, that’s about 768,000 tokens of idle context, and re-caching that much on a premium model such as Opus 5 runs somewhere around five to eight dollars. That’s real money spent to reload what the agent already had twenty minutes ago.

The fix is the same habit as above. Keep your working context small, and try not to leave a giant session parked when you step away. A smaller context is cheaper to carry and cheaper to rebuild.

4. Tell the agent what you actually want

When your request is fuzzy, the agent fills the gap by exploring, and every guess costs it a few file reads. Compare "the signup thing is broken" with "add email validation to the signup form and show a clear error when it fails." The second one tells the agent exactly what to build, so it doesn’t have to work that out first. A little more detail from you saves a lot of digging from it. Nobody has ever regretted being clear with a robot.

5. Match the model to the job

You don’t need the biggest model for everything. Plenty of everyday work (small edits, quick renames) runs just fine on a lighter, faster model. Reaching for the most powerful one every time, out of habit, is like hiring a full orchestra to play Happy Birthday. A good rule is to try the lighter model first (for example, Sonnet before Opus) and only move up when the task really asks for it.

6. Break giant files into smaller ones

If a single file runs three thousand lines, the agent usually has to load the whole thing to change one small part. Smaller, focused files mean it only pulls in what it needs. Your future self reading that code will quietly thank you as well.

7. Don’t paste huge logs or full dependency trees

When something breaks, paste the handful of lines that matter, not the entire stack trace with every warning stacked above it. The agent doesn’t need the whole novel to understand the plot. Trim first, then share.

Wrapping up

You don’t need new tools or a fancy setup to minimize your token costs: just tell the agent where to go, then clear away the clutter it’d otherwise have to read through. By doing that, you can get the same work done for a fraction of the tokens. 

^{Interested in cutting down on more of your AI costs? We offer an ROI-first AI strategy to help our clients make the most out of their solutions.}