The Time Coding Agents Spend Waiting for Approval

Once a team gets comfortable with Claude Code or Codex, people stop running one session at a time. One fixes a screen, one works on the API, one cleans up tests. Each runs in its own terminal.

That is where an unexpected wait appears.

Agents ask a person before they run certain commands, or when they are unsure how to proceed. That is a safety feature and it should stay. But the request shows up inside a terminal tab nobody is looking at. Until someone notices, the agent sits there doing nothing.

The work was supposed to be faster, yet overall progress is slower than expected. Often the cause is not the model at all. It is this time spent waiting for approval.

First: decide what a person checks

Put a human check on every command and the waiting grows. Allow everything automatically and operations that touch production or customer data go through too.

The first decision is where to draw that line. Operations that are safe to repeat, like running tests or a build, can be allowed automatically. Operations that are hard to undo, like deleting, deploying or sending data out, always get a person.

Claude Code and Codex both let you set permission rules. Once the line is decided and written down as rules, the number of requests that reach a person drops on its own.

Second: decide where the request arrives

Even with good rules, some requests still need a person. What matters is where they show up.

Inside a terminal tab, nobody notices unless they go and look. The request should land somewhere visible wherever you are on the screen, and you should be able to answer it right there. Removing the step of finding the right terminal and switching to it shortens the wait a lot.

Third: choose tools your company can approve

Tools that sit next to coding agents run right beside the code and the prompts. The conditions that make them easy for an IT team to approve are clear.

Your code stays on the Mac. Sessions, prompts and code go to no server anywhere.

Setting changes can be undone. If a tool edits the agent's settings files, it backs them up first and removes only what it added when you disconnect.

Answers mean the same thing. An answer sent from the tool is treated exactly like one typed in the terminal. Responsibility does not move, and that matters when the tool is reviewed.

The tool I built

To solve this for myself I built Kenko. A character from my manga lives in the MacBook notch and watches your Claude Code and Codex sessions.

While the agents work, she stays quiet. She only comes down when an agent asks for permission, asks a question, is waiting for input in the terminal, hits an error, or when your plan limit nears 80% and 95%.

Permission requests can be answered in the notch with Allow, Deny or Always. Questions can be answered with one of the options or your own text, and the answer goes back to the agent exactly as if you had typed it in the terminal. If you would rather handle it yourself, a button brings that terminal to the front.

It uses the hooks feature in Claude Code and Codex. Connecting backs up the settings files first, and disconnecting removes only what it added. The app talks to the hooks over a local socket on the Mac, so sessions, prompts and code never leave the Mac. There is no account and no analytics in the app.

It runs on macOS 14 or later and costs $0.99, once. Get it at kenkoisland.com.

Speed comes from designing the wait

I think what separates teams adopting agents is less the model and more how the handoff between people and agents is designed. What a person checks. Where the request arrives. Whether the tools can be approved internally. With those three in place, running several agents at once stops turning people into the bottleneck.

If your team is starting to run agents, tell me three things in the first meeting. The agents you use or are considering. The operations you think need a human check. And the step that takes longest when your company approves a new tool.

It is fine if you do not know all of them yet. We can start by sorting that out together.