Talk to me

Why does my coding agent hit its usage limit so fast?

A short learning video. A renovation company owner who used up her limit with one short question, and a software engineer who shows her what really goes out on each step.

Video. A software engineer shows a business owner why one short question used up her limit, and what to change. 2 min 55 s, two voices, quiet music. Read the article.

What this video shows

Nora asks one short question about a small bug, in a session open since Monday, and her limit is gone by lunch. Sam draws what goes out on each step: her question, the rules, the files and the test results, all sent again each time. Then he shows one real session, 48 minutes long, where each step grew from about forty thousand tokens to over three hundred thousand.

The second half is what to do. One task in one session, with a short note. A fresh start after a long break. Short rules and fewer tool servers. A smaller model for easy work, chosen at the start. Running several agents adds to this, as video 22, “Can I run several AI coding agents at once?” shows. The article has the real chart, each tool’s limits and a short check you can run today: read the article.

What you will learn

  • Why a short question can use so much of your limit.
  • Why one task in one session is the first thing to change.
  • What a long break and a model switch do to your usage.
  • Why a tool that shrinks everything can cost more.

Further reading

  • Anthropic, Claude Code docs: “Manage costs effectively” and “Prompt caching”. costs and prompt caching
  • OpenAI, Codex docs: Pricing (usage limits). learn.chatgpt.com
  • Liu and others, July 2026: “Agentic Coding in the Wild: Characterizing GitHub Copilot at Production Scale”. arXiv 2608.00101
  • Weinberger and Hozez, 2026: “Token Reduction Is Not Cost Reduction”. arXiv 2607.12161

If your team wants coding agents that do real work and stay within their limits, that is the kind of setup my team builds. Questions are welcome.

Transcript

0:03 Nora: Why does my coding agent hit its usage limit so fast?

0:09 Sam: Because each step sends the conversation again, not only your last message. So start a new session for each task. Keep what loads at the start small. And use a smaller model for easy work.

0:23 Nora: This morning I asked one short question about a small bug. By lunch, my limit was gone.

Sam: Was it a new session?

Nora: No. The same one since Monday.

0:34 Sam: Look at what goes out on each step. First, your question, with the rules it loaded. Then it reads a file, and sends all of that again, with the file. Then it runs a test, and sends everything again, with the result. Then the next file.

0:52 Nora: So my short question carried the whole week?

Sam: Yes. Claude Code’s docs say a one-line question still draws usage for the whole conversation.

1:02 Sam: Here is one real session from an agent team. It ran for forty-eight minutes. At the start, each step sent about forty thousand tokens. A token is a small piece of text. At the end, each step sent over three hundred thousand. Almost all of it was old conversation, read again. Only a tiny part was new answers.

1:28 Sam: So, one task, one session. When the task is done, write a short note. What you did, and what is left. Then clear the session. Clear costs nothing.

1:39 Nora: And when I come back after lunch?

Sam: The saved copy of the conversation lasts only a short time. After a long break, your next message processes everything again. So start fresh.

1:53 Sam: Next, look at what loads at the start. Your rules file goes out on every step. So keep it short. Turn off the tool servers you do not use.

2:04 Nora: And the model?

Sam: A smaller model for easy work. Choose it at the start, because a switch in the middle can lose the saved copy. Extra agents add their own requests too.

2:17 Nora: Can a tool just shrink everything for me?

Sam: Be careful. In one test, the strongest setup cut tool output by over a third, and the bill went up. Send less instead. Ask for the error lines, not the whole log.

2:32 Nora: Today I did four tasks, in four new sessions. I still have room.

Sam: Good. Once a week, open the usage report. It shows what used your limit.

2:42 Sam: Each step sends it all again. So one task, one session. Short rules. And the right model.

Have an app idea, a stuck app, or a daily job you want an AI agent to do?

Write three lines. I reply within one working day with a plain answer.