What's the last thing you fixed manually that you'll have to fix again next month?
For developer Rutger de Knijf, the answer was: almost everything. System maintenance, permission audits, tax document checks, trip routing - all solvable once, but requiring re-solving every few weeks. His fix: create a dedicated Claude Code agent for each recurring problem and let it handle execution while you review.
The recipe is simple. Create a folder. Run claude init. Write the agent's role and the problem it owns into a CLAUDE.md file. The agent executes, you review its output, and when the result needs work, the agent updates its own instructions for next time. Instructions tighten over multiple runs, and your review time shrinks.
De Knijf applies this across a wide range: a system maintenance agent that checks backups, firewall rules, DNS, and software updates every few days; a permissions agent that finds and deduplicates access files; a vacation planner that optimizes EV charging routes and generates language flashcard decks for each destination; a finances agent that validates tax documents.
The prerequisites are real. You need a flat-rate token plan - per-message pricing makes agent loops expensive fast. You need MCP server connections for any external data sources. And you need patience to iterate on the instructions until they're precise.
The key shift is in the human role. You're not fixing problems anymore - you're defining them precisely enough that an agent can own them. For most recurring tasks, that investment pays off after the second or third run.