Maintaining WordPress Sites With Agents, Not Cron Jobs

The worst part of running WordPress for clients is the monthly maintenance: plugin updates, broken links, orphaned media, slow queries. I used to do it manually and charge for it. Now an agent does the first pass and I review. Eight months in, here is the workflow that works.

The Weekly Maintenance Loop

Every Sunday night, a scheduled task triggers an agent with the client site's credentials and a playbook. The playbook is a markdown file listing what to check and what to do if something is wrong. The agent produces a report; I read it Monday morning; I approve or override fixes.

The Playbook Format

# MAINTENANCE.md
 
## Checks
- Plugin versions vs latest (core, Yoast, WooCommerce, contact forms)
- Core version
- PHP version
- Dead links in posts published last 90 days
- Uploads folder size vs last month
- Slowest 5 queries from last 24h (via Query Monitor)
 
## Auto-fix (agent does without asking)
- Minor plugin updates (x.y.Z)
- Broken internal links with matching slug elsewhere
 
## Propose-only (agent reports, I decide)
- Major plugin updates (X.0 or x.Y.0)
- Plugin deactivation / replacement
- PHP version changes

The propose-only section is the safety valve. Breaking changes never ship on autopilot.

The Rollback Requirement

Every auto-fix must be reversible in one command. Before any plugin update, the agent triggers a snapshot via UpdraftPlus or the hosting provider's backup. A fix I cannot undo is not an auto-fix, it is a gamble. I do not gamble on client sites.

What the Client Sees

A weekly email: what was checked, what was updated, what needs attention. Three sections, no jargon. The client pays the retainer because the site stays healthy without their involvement. They do not need to know an agent did the first pass.

The Unglamorous Truth

This is not flashy work. No one writes thinkpieces about plugin updates. But it is where agents earn their keep on real client sites. Small, repetitive, high-stakes-if-skipped work is exactly the shape that agent-assisted maintenance fits.

Scaling the Pattern

One playbook per site does not scale past a handful of clients. The next step, which I moved to in month seven, is a shared base playbook plus a per-client override file. The base gets the common checks. The override captures what is unique: an e-commerce client's stock sync, a membership site's subscription cleanup. The agent merges both at runtime. Adding a new client becomes a fifteen-minute override file, not a playbook rewrite.