Step 5 of 6
Watch progress from outside the agent
A finished task and a stalled one look the same in a quiet chat. Track progress in something the agent does not write: the files, the clock, or a dashboard.
What happened
Case pending
This step is written from repeated use, not from one recorded incident. A dated, first-person account will replace this block before the site leaves preview. Until then, treat the steps below as guidance that has held up, not as a verified story.
What it costs when you skip it
You start a long task, come back an hour later, and the chat looks exactly as it did. No error, no progress, no answer. The task stopped twenty minutes in. The hour is gone and the work was never going to arrive.
The same silence also hides a real success, which trains you to interrupt things that were fine. Either way you are guessing, and guessing is the expensive part.
Do this
- Write down the deliverable before it starts. A named file in a named folder, with a size you would accept. "Done" has to be checkable from outside.
- Ask for a running log. "Append one line to
logs/progress.mdafter each step, with the time." The log is a file, so you can watch it without asking. - Watch the file clock, not the chat. Check the modification time of the output folder. If nothing in the workspace changed for ten minutes during a job, assume it is stuck.
- Set a stall rule in advance. Ten minutes for interactive work, longer for jobs that legitimately run long. Write the threshold down so the decision is not made while you are annoyed.
- Interrupt with a status question, not a restart. "Status: which step, what is blocking, keep going or stop?" Restarting loses the work that did succeed.
- Use a board if you run several at once. One row per task, one progress value, updated by something other than the agent's own text. Brake Board is being built for exactly this.
| In the chat | What is actually true |
|---|---|
| "Still working on it" | Nothing has been written for twelve minutes. It is waiting on a permission prompt you never saw. |
| No message at all | The process exited. Usually silent, occasionally with an error above the fold. |
| A long pause after a tool call | It is retrying the same failing command on a loop. |
| "Done" | Sometimes a plan, not a result. Check the file exists and opens. |
Checklist
Common mistakes
- Asking "how is it going?". The answer is generated from the same context that is stuck. "Nearly done" is not evidence.
- Checking only at the end. A job that stalled early and one that ran the whole time end in the same place: no file. You only learn the difference from the log.
- One enormous task. Twenty steps in one request has one progress value: zero or done. Break it into five and you can see where it stopped.
- Restarting instead of asking. Restart plus lost work is the most expensive possible response to a stall.
What you need for this step
Brake Board is the tool for this step. It is in development and not open for testers yet; the waitlist page states exactly where it stands.
Join the Board waitlist