“AI slop” has been one of LinkedIn’s main subjects for months. The platform added a way to report low-effort AI content, says more than a million people have used it, and reports that flagged posts are now seeing far fewer views. This week one of its engineering leaders described how LinkedIn detects slop and trains its own agents to avoid it.

The backlash is fair. Much of what AI produces in public is volume without substance. It’s worth remembering, though, that the most useful AI work tends to happen where nobody posts about it.

Here’s one example from inside Calendrz, the calendar product I build. Every night at about 3 a.m. Central, an AI agent reads our production logs. Since late April it has opened 42 fix pull requests that we’ve reviewed and merged. Here’s how it works and what keeps it from producing slop of its own.

The job: read the logs nobody wants to read

Calendrz mirrors availability across many Google and Microsoft 365 calendars. That means a constant stream of OAuth refreshes, webhook notifications, API retries, payment events, and a long tail of internet scanners probing for WordPress admin pages we don’t have. The server logs record all of it, and almost none of it needs a human. A small part of it really does matter.

The nightly agent runs in GitHub Actions on Claude Code. Its instructions are deliberately narrow:

1. Read the last 24 hours of WARN and ERROR entries, including multi-line stack traces.
2. If the logs are clean, say so and stop. No branch, no PR, no “summary” for its own sake.
3. Classify every distinct finding. Is it a real defect, a noisy log that should be quieter, an expected upstream failure, or a bot probe?
4. Open a fix PR only if it has a high-confidence fix for every finding. Otherwise it opens a GitHub issue that describes the problem and leaves the fix to a human.
5. Add new bot-probe paths to a shared blocklist that the app refreshes hourly, without a redeploy.

A human reviews and merges every PR. The agent never deploys anything.

What it has actually caught

Over the past few weeks, nightly PRs have:

• Added retries for transient Google Calendar write failures, so a brief upstream error no longer leaves a calendar without its blocker.
• Fixed Microsoft token refreshes that were sending an unresolved redirect URI, a bug that only appeared in specific refresh paths.
• Stopped a cleanup job retrying dead calendar tokens over and over for accounts their owners had already disconnected.
• Fixed a race condition where contacts added to an email list in quick succession were being silently dropped.
• Marked an email-list call as failed when every contact was rejected, instead of reporting success.
• Turned scanner noise down to DEBUG (Log4Shell probes, XSS-laden Accept headers, path-traversal attempts), so real warnings stand out.

None of these would make a headline, and most customers would never have noticed them individually. Together, though, they’re the difference between a service that quietly gets more reliable and one that quietly degrades.

Why this doesn’t turn into slop

Slop is output optimised for looking like work. The design choices that keep this agent useful are the same ones that separate good AI content from bad:

It is allowed to do nothing

“Clean logs, exiting” is a first-class outcome. An agent that has to produce something every night will invent something. Ours doesn’t have to.

It has to show evidence

Every PR links the log lines that prompted it and explains the change. An unsupported claim gets closed on review. That’s the same standard LinkedIn’s readers are now applying to posts.

Confidence decides whether it fixes or reports

When the agent isn’t sure, it writes an issue rather than a fix. A clear “here’s what I saw, I don’t know why” beats a confident patch for the wrong problem.

The task is narrow

The agent isn’t told to “improve the codebase”. It’s told to read one log, for one day, against one checklist. Narrow scope keeps quality high.

A human approves every change

The agent proposes and a person merges. Reviewing takes minutes. Finding these issues by hand would take hours.

What we’d tell anyone tempted to automate

If your AI work is aimed outward, at posts, emails or articles, the slop backlash is a useful warning. AI does its best work on tasks that are too tedious for people to do consistently, with clear inputs and an answer you can check. Nightly log triage fits that description. So does checking five calendars before you accept a meeting, which is why Calendrz exists in the first place.

Use AI where it removes drudgery, and hold its output to the standard you’d apply to a colleague.

If you’re running something similar, or thinking about it, I’d like to compare notes.