Hitting AI Agent Usage Limits? Ten Fixes to Try Before Upgrading Your Plan

If Codex or Claude Code keeps hitting its usage limit, the fix is often not a bigger plan. Most token waste comes from habits: buying credits without checking the meter, letting an agent click through websites like a person, running every task on the most expensive settings, and dropping huge files into the chat. Tokens are the units AI models use to process text, and they are what subscription limits count. Below are ten changes, grouped into four areas, that help the same subscription get more work done.

Read the meter before you change anything

Start by finding out where your usage actually goes. Small businesses often run several agents without watching the meter, then upgrade to a pricier plan they did not need. Sometimes the real fix is to stop redundant agent runs or to use free resets that are already in the account.

In Codex, a free reset restores the 5-hour or weekly limit without spending credits, but each reset has an expiration date, so it is lost if you ignore it. A setting called “Allow Codex to use resets” lets the agent use an available reset on its own when it reaches a limit. Be far more careful with automatic reloads of paid credits: an agent stuck in a loop can drain a meaningful amount of money in minutes. Before you click to buy more, look at both the usage bar and the list of resets.

Claude Code appears to grant free resets much less often than Codex, which makes efficient routing more important there. Its weekly breakdown by product can be revealing. In one heavy user’s week, Claude Code accounted for 96% of usage and regular chats for 4%. For people working inside the Claude desktop app, the advice is to keep agent work in Claude Code rather than spreading it across chats or Cowork.

Stop agents from clicking through websites

When an agent logs into software and navigates web pages the way a person would, token use climbs and speed drops. The same work runs faster and lighter through a plugin, an MCP server or an API key. MCP, the Model Context Protocol, is a standard way for AI tools to connect to outside software and data. An API is a direct programmatic interface that a program can call without a browser.

The difference shows up clearly in real cases:

• A business that had operated for more than ten years asked an agent to audit its HubSpot account without an API key or MCP. The agent used up a full week of subscription limits in one day.
• Another business had adopted GoHighLevel on an agency’s recommendation. Its MCP server and API could not build or edit funnels and workflows, so the agent had to click through the interface, burning large amounts of tokens at a slow pace.
• A company with more than 60,000 files in version 1 of Basecamp could not connect agents at all. After migrating its data to Basecamp V5 with the vendor’s support team, Codex and Claude agents could audit operations and automate workflows directly. Staff kept a familiar tool while agents gained full read and write access.
• An attempt to have an agent upload a video to YouTube failed over to the browser because the API key only had read permission. A browser route can take 30 minutes for an action an API call finishes in 30 seconds, so it was worth stopping the agent and spending 20 minutes setting up upload permissions instead.

A simple audit helps. Ask your agent to research each core tool your business runs on and answer four questions:

1. Is there a plugin or connector?
2. Is there an MCP server?
3. Is there an API with read and write access?
4. Is there a newer version of the software with better access?

Have it also list which tasks can run through those routes and which would still need browser clicking. Then add a rule: if the agent starts using the browser for a tool that has an API or MCP route, it must stop and ask you. If a tool has no agent access and no plans for it, you face a choice between paying the ongoing token cost and switching tools. For established companies that need deep agent compatibility, HubSpot is the recommended CRM, or customer relationship management system; for smaller businesses, Notion is suggested as an agent-friendly option.

Move heavy daily workflows from plugins to APIs

Plugins are a fine start, but any workflow that runs every day or moves a lot of data belongs on a direct API. APIs bring speed, lower cost and field-level control, and they do not require a full engineering team. A side-by-side comparison of the HubSpot ChatGPT plugin and a HubSpot private-app API token makes the gap concrete:

An API can also fetch only the fields you need. A daily sales report that needs client name, stage and deal amount gets exactly those fields through the API, while the plugin loads every CRM property on every record. Direct tokens also handle custom properties, scheduled jobs that run unattended, and links to accounting systems such as BigTime.

Store API keys in a dedicated credential manager with permission scopes set in advance, and keep logs so you can trace which agent took which action if something goes wrong. Turn off connectors you do not use. Their tool descriptions are added to every prompt, taking up space in the context window, the amount of text a model can consider at once. Direct APIs also keep you free to switch models and tools, moving between ChatGPT Work, Codex and Claude Code or using orchestration tools that swap the underlying model. Browser logins and platform-specific plugins tend to lock you into one environment.

▲ Browser clicking versus a direct API connection

Match effort, model and speed to the job

Reasoning effort controls how much a model thinks before answering. For everyday work, leave it on Medium. Claude Opus 5.5 on Medium can handle roughly 90% to 95% of regular tasks, and spending maximum effort on a follow-up email quickly empties a daily budget. Higher effort without clear boundary rules may simply help an agent make wrong decisions faster.

The base model matters as much as the effort setting:

Speed settings deserve a second look too. In Codex, a lightning-bolt toggle next to GPT-6 Astra delivers 1.5x speed but burns through the quota faster. Background agents usually work while people do something else, so paying extra for speed rarely helps. Leave it off for unattended tasks.

Keep AGENTS.md and CLAUDE.md short

AGENTS.md (for Codex and ChatGPT Work) and CLAUDE.md (for Claude Code) are the onboarding handbooks agents read at the start of their work. Files written in the 2024 style spent hundreds of tokens explaining how to write an email or listing every folder path. Current models already know standard formatting and can explore directories on their own, so the handbook should hold only knowledge that is specific to your organization:

• Who owns which area of work
• What the agent is allowed to do
• Which system is the source of truth when data conflicts, for example QuickBooks rather than HubSpot for payment status
• When the agent must stop and ask a person
• A changelog of past errors and how they were fixed

You can ask an agent to investigate your connected systems and files and draft the file for you. Review it every three to six months, because stale instructions can confuse newer models and hold back their performance.

Put files in one place with predictable names

Many small businesses keep documents in five to seven places: Dropbox, Google Drive, iCloud, SharePoint, local desktops and former employees’ accounts. An agent that does not know where the authoritative copy lives has to search all of them, spending tokens and risking an outdated file. In one case, an agent looking for last year’s pricing sheet found three files named “pricing final” and emailed an old version to a prospect.

A cleanup follows six steps:

1. Pick one storage system. If you already use Google Workspace, consolidating in Google Drive works well.
2. Build folders around how the business actually operates, such as company-wide folders for brand, marketing, finance, procedures and templates, and client folders for proposals, contracts, deliverables, meetings and archives.
3. Adopt a naming convention. For client documents, use date, client, document type and version in that order, with the date first so files sort chronologically. Leave dates off evergreen assets such as logos and core procedures.
4. Remove duplicates, and replace labels like “FINAL” or “USE THIS” with version numbers so agents do not have to guess which file is current.
5. Archive material that is no longer in use.
6. Check that a new team member could understand the structure.

Begin with a read-only audit: ask the agent to catalog storage locations, file types and duplicates and to propose a cleanup plan, without moving or deleting anything until you approve. When the move is done, record the source-of-truth locations, naming rules and folder layout in AGENTS.md. Consistent names act as a small index that lets an agent pick the right document without opening it.

▲ Files consolidated into one structured store

Turn great work into a standard

When an agent produces an excellent deliverable, capture the prompt, structure and criteria right away as a template, a skill or a checklist. A skill here is a saved set of instructions the agent can reuse. Without that step, an owner can spend an hour every month reteaching fonts, layout and content rules from scratch. Codifying one good result a week would add up to about 250 reusable assets per team member in a year.

Good candidates span six areas: sales (pitch decks, pricing sheets, follow-up emails), client delivery (onboarding packets, quarterly reports), marketing (landing pages, newsletters, lead magnets), social media (LinkedIn posts, carousels), operations (standard procedures, payment reminders) and people (job postings, onboarding checklists). For an HR compliance checklist, for example, save the approved version as a locked master template, then create a skill that lists required inputs, things the agent must not do, checks to run before handoff and a review date six months out, and register it in AGENTS.md.

Test every new skill before you rely on it. For a slide design standard, ask the agent to build a fresh three-slide sample using only the style guide, fix what is off, and then save the standard. Proven sales decks or signed agreements are good raw material, so you do not have to start from a blank page.

Point to files instead of uploading them

Dragging a 150-page contract, an hour-long meeting transcript or a large spreadsheet into a chat keeps that content in the conversation, and it is processed again on every turn. Upload a whole contract to check one cancellation clause, and every later question carries those 150 pages with it.

Tell the agent where the file is and what part it needs:

• Instead of uploading a 200-page contract, ask it to review the addendums in the client’s contracts folder in Google Drive and flag the pricing terms.
• Instead of a full transcript, ask it to pull the sales objections from Tuesday’s call transcript in the meetings folder.
• Instead of pasting a 5,000-row dataset, ask it to find the top ten clients by revenue in the Q3 spreadsheet in the finance folder.

A file in a folder costs nothing until the agent opens it, while a file uploaded to the chat keeps adding cost for the rest of the conversation. You can also add a rule to AGENTS.md telling agents not to ask for full uploads of large documents.

What to do this week

Usage limits seem to leak through work habits more often than through plan size. A practical order of operations:

1. Run /status in Codex or /usage in Claude Code, note remaining limits and reset expiration dates, and review any automatic credit reloads.
2. Audit your core tools for plugins, MCP servers, APIs and newer versions, and move browser-based tasks onto those routes.
3. Turn off unused connectors, default to Medium reasoning, send simple tasks to lighter models, and switch off speed boosts for background work.
4. Trim AGENTS.md or CLAUDE.md to organization-specific rules, and consolidate your files under one naming scheme.
5. Save great outputs as tested skills, and point agents to file locations instead of uploading large files.