❯ 5 terminal commands I run on my AI agents every week

You do not need to be a sysadmin to get value from running an AI agent. But the people who run agents well seem to end up in the same place: a terminal window. This is the plain-English version of why, and the five checks that let a non-terminal person trust an agent without living in a shell.

← back to articles

I don't have an agent dashboard. I have a terminal. Every automation I run — an AI agent that files receipts, writes the weekly digest, pokes the CRM — gets checked the same way I check a webserver: on the command line. Here are the five things I actually run, and why each one earned its spot.

1. journalctl — did it run, and did it stop?

The fastest question for any agent: did it run at all, and did it stop cleanly or die mid-loop?

journalctl -u my-agent --since "yesterday" | tail -40

I'm looking for three things: a start, no killed/OOM lines, and a clean finished/exit. A job that simply never ran is louder than one that errored — because nobody configured a failure alert for a job that silently didn't start.

2. tail the log, then grep for the exceptions line

Logs are 99% noise. I only want the interesting 1%:

tail -f /var/log/agent.log | grep -Ei 'error|retry [0-9]|gave up|unexpected'

Two signals I care about: a retry count climbing (the loop is stuck) and "gave up" (the loop stopped, but did it tell anyone?). If the answer to the second is no, add it to the failure path.

3. ps — is it actually working, or just running?

A process can be alive and useless. I check what the agent is doing right now:

ps -o pid,etime,%cpu,comm --sort=-%cpu | head -8

Constant high CPU with no output = a spin. Eternity with 0% CPU = it's waiting for something that will never arrive. Both are different failures.

4. curl the health endpoint — not the happy path, the "3am" question

A real agent has a /health that answers "can you still reach your tools?":

curl -sf http://localhost:8080/health && echo OK || echo DOWN

I run it on the CI box and from a cron. If the agent can't reach its tools at 3am, I want a page, not a discovery next Tuesday.

5. tail the audit — what did it actually do?

The terminal's best feature is that it doesn't sugarcoat. The audit log tells truth the dashboard hides:

grep -E 'action|tool|approve' /var/log/agent-audit.log | tail -25

I'm checking one thing: did it act within the classes I approved, or did it get autonomy I never granted? That one grep has caught more than every dashboard widget combined.

The pattern

All five reduce to the same instinct: agents are systems, so check them like systems — not like magic. They're not special. They boot, log, spin, fail, and need a human to notice. The terminal is just the honest place to look.

That's also the correct bar for a managed provider: if they can't hand you a log line, a retry cap, a health endpoint, and an audit trail, then their dashboard is hiding the same thing mine was.

Claw Way runs agents the way you'd want to run them in a terminal — loops capped, actions logged, health monitored, and an audit trail you can actually read. If you'd rather get the 3am page instead of sending it, we run them for you.

by Jonas Reyes — the builder's desk, Side Quest Studios

AI-assisted, curated for Side Quest Studios.

References

1. journalctl — systemd manual (the service log)
2. grep — manual (filtering log noise)
3. curl — manual (the health check)
4. ps — manual (is it working or just running)
5. 5 terminal commands I run on my AI agents every week (this post)

That is the whole trick: agents are systems, and a terminal is where systems tell the truth. You do not have to live there. But the next time a dashboard looks fine and you still have that feeling, open a terminal and run one command. The dark side has good uptime.