Last week I ran some experiments on a side project. An agent gets a repository with one bug in it and a shell. It runs the tests, or the linter, or pulls the log of the failed pipeline, reads what comes back and fixes the bug. I run each task a few times and count tokens. Everything is in the repository on GitHub.

The pipeline log is where it got silly. One run read the CI log the way the tool prints it and spent more than ten times the tokens of a run that read only the lines with errors. Same bug, same fix. The difference was the prefix. Every line starts with the job name, the step name and a timestamp, about seventy-five bytes before anything useful. In a browser nobody reads that. The agent paid for all of it.

Then I pasted the transcripts into a chat and asked a model what the agent had read. So I paid to read the output, and then paid again to read the reading. Several rounds of that. By Friday the analysis had cost more than the experiments.

It isn’t only CI logs. Test runs, plans, scans and builds all print for someone sitting at a terminal. Colors, a progress bar, a version banner, and two hundred passing tests listed one per line around the eight that failed. Someone on a public issue tracker counted one test run. About a hundred and fifty lines, and eight of them mattered. The issue was closed as “not planned”.

My first reaction is to cut. tail, a byte cap, the last hundred lines. Every time I did that, sooner or later the failing line was above the cut.

What I’d like is to run a command and get only what I’d act on. One line per failure or change. No color. The same bytes for the same input, so I can diff two runs. A line at the end saying what was left out and how to get it back. And never a value the tool knows is secret, because a byte cap has no idea which ones those are.

I also want the tool to decide what I need, which is the exact thing I complain about in dashboards other people built. I haven’t figured that one out.

Two test runners I found already switch to a quieter reporter when they see an agent in the environment variables. Small thing. It’s the first time I’d seen a tool treat the agent as a reader.

I could spend a week of API bills on this because the money was there and nobody was waiting on me. Someone on a free tier gets the same hundred and fifty lines in a smaller context window, and when the window fills, the agent forgets how the task started.

I don’t know where the fix belongs. Maybe the tools, one reporter at a time. Maybe the agent harness, which already knows how much room the model has. Maybe something in the pipe that translates, and that may not survive the year the tools catch up. I also deleted four of the formats I’d built, because the numbers said they saved nothing. The agent was fine with what the tool printed. So I’m not even sure the short version is always what it needs.