One coding-agent user says 40% of their messages over three weeks went to overhead rather than real work. That is a personal result, not an industry-wide rate—but the breakdown raises a useful question: how much of your time with coding agents goes to directing, correcting and checking them?
What the three-week analysis found
In a DEV Community article, Toruk Makto describes reviewing messages sent to several coding agents, mainly Claude Code and Kimi, and sometimes Cursor and Copilot. The author exported three weeks of messages and labeled each by purpose. They report that keyword searches produced inaccurate counts, so they did not use those counts for the final analysis.
As an Amazon Associate I earn from qualifying purchases.
More than half of the apparent user messages in the logs came from scripts and test harnesses running under the author’s usual configuration. After removing that automated traffic, the author counted 2,116 messages they considered their own—about 96 per day—and reported this distribution:
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problems| Share reported | Message purpose |
|---|---|
| 55% | Real work: new tasks, questions and decisions |
| 13% | Correcting the agent, such as fixing a misunderstood task, drift or an unrequested model or scope change |
| 9.5% | Asking for progress |
| 6% | Manually carrying information between agents or chats |
| 4% | Continuation prompts such as “go,” “yes” or “continue” |
| 4% | Asking for an explanation in simpler English |
| 3% | Repeating a rule already given |
| 5% | Other, including slash commands and fragments |
The author summed the non-work categories and described the result as: “So 40% of my typing is overhead.” The percentages are the author’s reported labels and are rounded; they should be read as a description of this one workflow, not as a precise benchmark for coding-agent users generally.
#1 Best Overall
Where the overhead came from
Corrections, especially during UI work
At 13%, corrections were the largest named overhead category. The author identifies UI work—correcting one screenshot at a time—as the biggest single source of those corrections. The account does not establish that UI work is generally more error-prone with agents; it shows where this author’s corrections clustered.
Progress checks during silent runs
The author says they asked for progress 200 times. More than half of those requests came in bursts within the same hour, while long-running tasks finished silently. That pattern suggests a practical distinction: messages spent checking status may reflect a visibility problem in a particular workflow, not necessarily a problem with the task itself.
Relaying context and repeating rules
The author describes passing reports manually between two agents 33 times on their worst day and repeating instructions across agents. If context or rules do not carry between tools or sessions, the user becomes the integration layer, spending messages on transfer rather than new work.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Unclear costs and small continuation prompts
The author also worried about agents starting costly runs without first explaining what they might cost. Short prompts such as “continue” or “go” formed another reported category. The article does not quantify the financial cost of those runs or show that a particular tool avoids them.
What the 40% figure does—and does not—tell you
This is a self-analysis by one author, with no representative sample, comparison group or independent validation of the message labels. Its title’s “Is it the same for you?” is a question to readers, not evidence that other users have the same overhead rate.
The result is still useful as a prompt for examining your own workflow. The author used several agents, but the account does not compare Claude Code, Kimi, Cursor or Copilot, nor demonstrate that any one of them resolves the issues described. A message count also does not measure time, effort or value: a brief correction and a detailed task may each count as one message while demanding very different amounts of work.
Rank #4
How to compare your experience
If you want to answer the question for your own work, track a representative period and classify messages by purpose. Keep automated traffic separate from your own prompts, and decide in advance how to label ambiguous messages. The original analysis reports that keyword searches miscounted messages; manual classification was the author’s chosen approach, not a validated universal method.
Useful categories to track include:
- New tasks, questions and decisions
- Corrections or scope changes
- Progress checks
- Context relayed between tools or chats
- Continuation prompts
- Requests for simpler explanations
- Repeated instructions
- Other messages, including commands
Message counts can reveal patterns, but add time or effort if you want to know which pattern costs you most. For example, frequent progress checks may take little time individually, while a small number of substantial corrections may consume much more. Those are questions to measure in your own workflow, not conclusions established by the author’s counts.
Best Value
Questions to ask about your workflow
Rather than assume a specific product will fix the problem, compare workflows by asking:
- Can you see meaningful progress during long runs, or do you repeatedly have to ask for status?
- Do important instructions persist across agents and sessions, or must you repeat them?
- How often do you relay information manually between tools or chats?
- Which tasks generate the most corrections, and are the corrections concentrated in work such as UI changes?
- Does the agent make the likely cost of a long run clear before starting?
Do you see the same problems, or is your overhead somewhere else? Which one costs you the most? Have you found a way to reduce progress polling or keep rules from failing to carry over between tools?
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
