I spent the last year wiring AI agents into small, boring workflows: email triage, deploy checks, content pipelines. Most advice online is about building agents. Almost nobody talks about operating them week after week. Here is what actually mattered.
1. The brief beats the prompt
Every reliable agent run in my setup starts from a task file: goal, constraints, stop conditions, and where to write results. Prompts evolved constantly; the brief format barely changed in a year. When a run goes sideways, ninety percent of the time the brief was ambiguous and the agent guessed. Fix the brief, not the temperature.
2. Agents need an audit trail more than they need memory
Fancy memory systems kept disappointing me. A plain append-only ledger — CSV or markdown, one line per action — outperformed all of them. The agent reads its own ledger at start, humans grep it when suspicious, and git history doubles as a timeline. Boring wins.
3. Give every agent its own blast radius
Separate browser profile, separate tokens, separate working directory. The first time an agent wiped the wrong folder, it was my folder because I was lazy about isolation. Never again. Cleanup after a bad run should be rm -rf sandbox, not an incident report.
4. Verification is the feature
An agent that claims success is worthless; an agent that proves it is priceless. Every task in my system ends with an independent check: curl the page, count the rows, diff the output. If the check cannot be automated, the task is not ready for an agent yet.
I write longer, less dev-flavored versions of these notes — aimed at people who want to use agents without becoming engineers — at arsentev.ai, including a free hands-on course that goes from AI user to AI builder in 11 modules.
What is the least glamorous thing you have automated with an agent? The unglamorous ones are the keepers.
Top comments (0)