In a three-week review of messages sent to coding agents, DEV Community author Toruk Makto classified 40% of 2,116 messages as overhead rather than real work. That is one person’s result—not an industry-wide rate—and the author’s examples point to several different sources of friction: correcting agents, checking progress, carrying context between chats, and repeating instructions.
What the 40% figure measures
Makto says they reviewed three weeks of messages sent while using several coding agents in parallel, mainly Claude Code and Kimi, and sometimes Cursor and Copilot. They removed automated traffic before counting 2,116 messages they considered their own—about 96 per day—and labeled each by purpose. The author says keyword searches produced inaccurate counts, so the final breakdown came from reviewing and labeling the messages rather than relying on keyword matches.
More than half of the apparent “user messages” in the logs were scripts and test harnesses running under the author’s usual setup. Those automated messages were excluded from the 2,116-message total. This distinction matters: the reported percentages describe the author’s labeled human messages after that cleanup, not every entry in the raw logs.
| Share of the author’s 2,116 messages | Category |
|---|---|
| 55% | Real work: new tasks, questions, and decisions |
| 13% | Corrections: wrong task, drift, or an unasked-for change in model or scope |
| 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’s headline summary is: “So 40% of my typing is overhead.” The categories are the author’s own labels, and the figures should be read as a personal accounting, not as a measured rate for coding-agent users generally.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
Where the overhead came from
Correcting work that drifted
Corrections were the largest individual overhead category, at 13% of the author’s messages. The author says UI work was the biggest single source of these corrections: they would review a screenshot, identify an issue, and correct it one screenshot at a time. The account does not establish that UI work is generally the most correction-prone kind of agent task; it describes this author’s workflow.
Checking whether a run had finished
The author says they asked for progress 200 times. More than half of those requests came in bursts within the same hour, while long runs finished silently. That suggests a specific kind of friction in this workflow: the user was unsure whether work was still underway or needed attention, so they kept checking.
Relaying context and repeating rules
When information did not move between agents or chats, the author manually passed it along. On their worst day, they relayed reports between two agents 33 times. They also found themselves restating rules already given. The article names several tools in use, but does not compare their context handling or show that one tool would eliminate these tasks.
Sending prompts that did not add much
Continuation prompts such as “go,” “yes,” and “continue” accounted for 4% of the author’s messages; requests for simpler explanations accounted for another 4%. These are separate categories in the author’s breakdown. The account identifies them as overhead but does not measure how much time each took or whether they could have been avoided.
What this does—and does not—say about coding agents
The result is useful as a prompt to inspect your own workflow, but it cannot answer how much overhead coding-agent users typically have. The article offers no representative sample, comparison group, or independent validation of the labels. Its “Is it the same for you?” is an invitation to compare experiences, not evidence that 40% is a common rate.
Nor does the account show that Claude Code, Kimi, Cursor, or Copilot caused or solved any particular category of overhead. They are tools the author says they used, not a controlled comparison. The author also raises concern about agents starting costly runs without first explaining what they might cost, but reports no measured spending or comparison of cost controls.
Rank #4
Questions to use when examining your own workflow
If you want to see whether your typing is serving the work or managing the agent, the author’s categories offer a practical way to review a sample of your own messages. Separate automated traffic from messages you actually sent, then label each message by its purpose. The result will depend on your tools, tasks, and labeling choices; it is a personal diagnostic, not a benchmark.
- How often do you correct a wrong task, drift, or an unrequested change in model or scope?
- Do you check for progress because a run’s status is unclear?
- How much context do you manually copy between agents or separate chats?
- Do instructions persist where you need them, or do you repeat the same rules?
- How often do you need to give a short continuation prompt before work proceeds?
- Before a long run begins, do you have enough visibility into what it may cost?
Which one costs you the most? Do you see the same problems, or is your overhead somewhere else? The author’s account leaves those questions open for other users to answer from their own experience.
Quick Recap
Best Value
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.
