Talk to me

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

A coding agent sends the whole conversation again on almost every step. One real session, measured request by request, the limits of Claude Code, Cursor and Codex, and what to change first.

A row of paper stacks growing taller, with a tiny one-line question on top of the last one: each step sends the whole conversation again.
Figure 1. Each step sends everything before it again. Your one-line question rides on top of the whole stack.

Because a coding agent sends the conversation again on almost every step, and it takes many steps for each thing you ask. Your one-line question also carries everything before it. So start a new session for each task, keep what loads at the start small, and use a smaller model for easy work.

What one real coding agent session used, step by step

Tokens sent in each request of one sessionEach dot is one request to the model. The line only goes up.

Tokens sent per request in one real coding agent sessionOne session, Opus 5.5, 22 October 2026, 48 minutes, 112 requests. The first request sent 41,385 tokens, request 25 sent 186,598, request 57 sent 271,431, request 89 sent 304,582 and the last, request 112, sent 327,812. 050k100k150k200k250k300k350k 1255075100112Request number Request 1: 41,385 tokens sentRequest 9: 92,829 tokens sentRequest 17: 166,357 tokens sentRequest 25: 186,598 tokens sentRequest 33: 212,688 tokens sentRequest 41: 236,489 tokens sentRequest 49: 254,455 tokens sentRequest 57: 271,431 tokens sentRequest 65: 284,938 tokens sentRequest 73: 290,738 tokens sentRequest 81: 297,908 tokens sentRequest 89: 304,582 tokens sentRequest 97: 309,189 tokens sentRequest 105: 314,216 tokens sentRequest 112: 327,812 tokens sent 41,385327,812

One real session, Opus 5.5, 22 Oct 2026, 48 minutes. My team’s own logs (token counts only, no message text).

See the numbers
RequestTokens sent
141,385
992,829
17166,357
25186,598
33212,688
41236,489
49254,455
57271,431
65284,938
73290,738
81297,908
89304,582
97309,189
105314,216
112327,812
Figure 2. One real coding agent session: tokens sent per request rise from 41,385 at the first request to 327,812 at request 112.

I run a team of agents every day in my Mac app, VShell. I took the token counts from one working session in our logs: 22 October 2026, Opus 5.5, 48 minutes, 112 requests. A token is a small piece of text, about three or four letters.

The first request carried 41,385 tokens. The last carried 327,812. The session went through 27.2 million tokens in total, and 98.6% of them were old conversation being read again from the cache (a short-term store that makes repeat text cheap). The model wrote only 90,106 tokens.

Cached text costs a twentieth of the normal input price on this model. It still came to about 56% of this session’s cost at API list prices. That share is my estimate, because Anthropic does not publish how a plan limit weighs each kind of token.

Claude Code’s docs say “a one-line question in a session that has been open all day still draws usage for the whole conversation”. Across all 1,087 sessions in these logs, cache reads were 92.3% of all tokens.

Video, 2 min 55 s, two voices, quiet music. 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. Open the video page.

What is inside one coding agent request

What one request carries, on averageShare of input tokens, GitHub Copilot agent sessions

  • Conversation history 48%
  • Tool calls and results 28%
  • System prompt 14%
  • Other 10%
  • 68,000tokens inmedian request
  • 247tokens outmedian request

Source: Liu et al., arXiv 2608.00101. GitHub Copilot, first week of June 2026, 13.5 million sessions. Output leaves out hidden thinking.

Figure 3. One coding agent request: conversation history 48%, tool calls and results 28%, system prompt 14%, other 10%. The median request sends 68,000 tokens and gets 247 back.

Microsoft researchers measured 13.5 million GitHub Copilot sessions in one week of June 2026. The median request sent 68,000 tokens and got 247 back, not counting hidden thinking. In another 2026 study, runs on the same task differed by up to 30 times in tokens.

What the usage limit is in Claude Code, Cursor and Codex

ToolThe limitWhere to see it
Claude CodePro and Max: a five-hour limit and a weekly limit. Shared with Claude chat./usage, /status, /context
CodexPlus: estimated 15 to 160 local messages per five hours on GPT-6.1 Sol, 350 to 3,000 on GPT-6 Luna. Weekly limits “may also apply”. Pro has no five-hour limit./status, usage dashboard
CursorA pool that refreshes weekly, shared with the Grok apps. “Different models draw from the pool at different rates”.Editor settings, Usage dashboard
Figure 4. The usage limit of Claude Code, Codex and Cursor, and where to see it, from each tool’s docs, read 3 November 2026.

All read on 3 November 2026. None of the three gives the limit as a token number. OpenAI’s range of 15 to 160 messages shows why: its docs say extended sessions “will use significantly more per message”. The Cursor pages I could read do not explain re-sending, but they price cache reads separately for every model.

Six things that drain a coding agent usage limit

What drains itHow to see itWhat to do
One long session for many tasks/context keeps climbing/clear, or a new chat, for each new task
Coming back after a breakCache misses in /usageStart fresh instead of continuing
A big model on small workSuggestions in /usage/model and /effort at the start
Heavy rules and MCP servers/context, /mcpTrim the file, turn off unused servers
Huge tool outputA long log inside the chatReturn only the lines you need
Agents you forgot/usage, by subagentStop them, check scheduled tasks
Figure 5. Six things that drain a coding agent usage limit, how to see each one and what to do about it. Commands are for Claude Code.

Commands are for Claude Code. In Codex, /status and /compact do the closest job.

Start a new session for each task

The same work, one session or threeTotal tokens sent, same requests, same growth per request

  1. One session27.1 M
  2. Same work in three sessions estimate10.7 M

Arithmetic on the same real session. A fresh session may need to read some files again.

Figure 6. One long session sent 27.1 million tokens; the same work split into three sessions would send an estimated 10.7 million.

I reran that session’s numbers as three shorter sessions, with the same growth per request. Total input fell from 27.1 million tokens to 10.7 million. A fresh session may need to read some files again, so treat 60% as the ceiling of the saving.

In my team each member takes one assignment in one session, writes a short report (what I did, the result, what is left), and exits. The next session reads the report, not the whole history.

/clear costs nothing. /compact reads the whole conversation to summarise it, so “compacting a large context is itself a large request”. Use it only to continue the same task. In the Copilot data, the 7.8% of sessions with a compaction held 44.2% of all tokens.

Why a break or a model switch uses more of your limit

The cache has a short life. On a Claude subscription, the main conversation gets one hour, and subagents get five minutes. On an API key or usage credits the main conversation also gets five minutes. After that, your next message “reprocesses your full context”. In Copilot, cache hits averaged 90% within a turn, 55% across turns and 8% after a model switch.

So pick the model when the session starts. My hands-on agents run one step below the top model. Anthropic says Sonnet “handles most coding tasks well”. Once you hit a Claude Code limit, /model will not bring access back, because the window is shared across all models.

Extra agents add their own requests, and an agent team in plan mode uses about 7 times the tokens of a normal session. My article on running several agents starts with two for this reason.

Keep what loads on every request small

Everything loaded at the start is sent again on every step. A rule of 1,000 tokens in a 112-request session is sent 112 times. My sessions start at about 50,000 tokens before any work, because they load a team brief and tools.

Claude Code’s docs say to keep CLAUDE.md under 200 lines. Cursor says to keep rules under 500 lines. An ETH Zurich study found context files raise cost by over 20% on average, so write only what the agent cannot guess. My article on a good CLAUDE.md shows how. OpenAI says “Every MCP server adds more context to your messages”; Claude Code now defers tool definitions by default.

Be careful with compression tools. In a test of 2,848 Claude Code runs, the most aggressive setup cut tool output tokens by 38.4% and raised the billed cost by 6.8%. Tool output was a small part of those short tasks, and agents with less to read took extra steps to look again. Long sessions may differ. Cut at the source: Claude Code’s docs show a hook that greps a 10,000-line log, taking it from tens of thousands of tokens to hundreds. And give a clear target, because “improve this codebase” makes the agent scan everything.

Usage limits in 2025, 2026 and a guess for 2027

  1. Weekly limits28 Aug: weekly limits on Claude Pro and Max.
  2. Bigger windows, raised limits1 million token window on default models. 22 Sep: five-hour limits raised.
  3. Both growLimits grow. Agents run longer.
Figure 7. Usage limits over three years: weekly limits in 2025, larger context windows and raised five-hour limits in 2026, and a marked guess for 2027. The dashed box is my guess.

How to check why your coding agent used its limit

Limit gone fast?

Open /usage or /status

  • One old sessionNew session per task
  • High number in /context at the startTrim rules and servers
  • Many agents or scheduled tasksStop what you do not need

Still high? Smaller model, lower effort

Figure 8. A flow to check why a coding agent used its limit: open the usage report, then fix old sessions, heavy start-up files or extra agents, then try a smaller model.
  1. Run /usage and press w for the last 7 days. Claude Code lists what drove your usage by skill, subagent, plugin and MCP server. It counts this machine only.
  2. Run /context in a fresh session. That number goes out on every step.
  3. Fix the biggest one and watch for a week, since one task can differ widely from the next.

Setting this up for a team’s agents is part of what my team does, and here is how we work. Which tool fits which job is in my comparison of Claude Code, Cursor and Codex.

Why did my limit go? Three checks

  • 1. /usage, press w: what drove the last 7 days
  • 2. /context in a fresh session: what goes out on every step
  • 3. Fix the biggest one, watch for a week

One task, one session.

Figure 9. Three steps to find what used up your coding agent’s limit, each with a tick box. Print it or keep it next to your screen.

Questions people ask

Does chatting in the Claude app use my Claude Code limit?

Yes. “These limits are shared across Claude and Claude Code”. Codex works the same way with ChatGPT Work.

Why did my usage go up when I was doing nothing?

A scheduled task “fires on its interval even while the session is idle, sending your full context each time”. Active teammates also keep using tokens until they exit. Usage on another device also counts, and the breakdown in /usage does not show it.

Will a bigger plan fix it?

It buys more room, and the same session still grows the same way. Fix the habit first, then judge the plan. Claude’s usage credits, Codex credits and Cursor top-ups let you keep working past a limit.

If you want coding agents set up for your team, working side by side with real checks, see how I can help or tell me about your project.

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.

Sources. My own session logs (22 Oct 2026; all sessions to 2 Nov 2026; counts only). Claude Code docs: Manage costs, Prompt caching, Context window, What’s new. Claude Help Center: Using Claude Code with your Pro or Max plan. Claude API docs: Prompt caching. OpenAI: Codex Pricing, Best practices. Cursor docs: Models, Rules (all read 3 Nov 2026). Anthropic, Opus 5.5 news (22 Sep 2026). TechCrunch (28 Jul 2025). Liu et al., arXiv 2608.00101 (30 Jul 2026). Bai et al., arXiv 2604.22750 (v3, 2 Oct 2026). Weinberger and Hozez, arXiv 2607.12161 (v5, 12 Aug 2026). Gloaguen et al., arXiv 2602.11988 (v3, 29 Sep 2026).