Automated Review★

Automated review is the process of using an LLM to evaluate code modifications. Unlike deterministic Automated Checks (which test if code compiles), automated review is non-deterministic: it evaluates code design, security risks, documentation compliance, and naming conventions at machine cost.

You set up automated review whenever you:

• Use a PR review bot that posts comments on GitHub diffs.
• Configure a secondary model in your agent loop to check if the generated code matches your project specs.
• Enforce a separate security audit prompt before merging code.

The Power of Context Separation

The defining rule of automated review is Context Separation: never ask the agent that wrote the code to review its own work.

If an agent has spent 20 turns writing a complex function:

1. Confirmation Bias: The agent's context window is stuffed with its own reasoning logs. If you ask it "Is this correct?", it will re-read its own assumptions and confidently state that the code is flawless.
2. The Solution: Spawn a secondary model with a completely clean context window. Feed it only the final git diff and the original spec. Because the reviewer has no attachment to the journey, it looks at the diff like a stranger, immediately catching logical gaps, missing edge cases, or redundant helper files.

Working Agent (Context Bloated) ──► Generates Diff ──► Reviewer Agent (Clean Context) ──► Unbiased Audit Report

Field Applications

In advanced agent pipelines, developers use different model tiers for review:

• The Writer: A fast, low-cost model (e.g. Claude Haiku) writes the code.
• The Judge: A larger, reasoning-heavy model (e.g. Claude Sonnet) runs the automated review on the git diff, checking for security vulnerabilities before the PR is created.