Stopping Code Slop with Continuous AI

Over 25 days, Continue generated close to 700,000 lines of code in parallel with Claude Code, but mixed in was slop: duplicate logic, messy abstractions, unnecessary comments, security vulnerabilities, and inefficient DB queries.

Instead of having humans catch these issues one by one, the Continue team built an Anti-Slop agent that embodies the team's standards and automatically reviews every PR in CI. The agent examines the changed files, and if it finds slop patterns, it fixes only what's needed in at most one file. If the code is already clean, it does nothing.

The core prompt follows the rule "clean up parts of AI-generated code patterns that hurt readability and maintainability," restricting the agent to look only at the PR's changed files and make targeted fixes. On top of this, the team adds its own slop definitions, such as TypeORM's n+1 queries and duplicate methods that arise from failing to find shared packages.

This approach works because, while off-the-shelf agents are strong at bug detection, they struggle to internalize all of a team's implicit code-review standards at once. Placing a fresh-context agent as a separate review step is more reliable and cheaper, and it lets humans review a narrowly scoped issue at a time as well.

What Continue calls Continuous AI refers to these signal-driven automated agents. When a PR opens, it gets reviewed; when an alert comes in, it gets triaged; and separate agents continuously handle security checks, performance audits, documentation updates, and dependency reviews. The goal isn't to replace developers, but to systematize work that was already being repeated by humans, freeing people to focus on more important judgment calls.