AI Agents for Code Review and Knowledge Sharing

Code review is one of the most knowledge-dense parts of engineering. Every PR is a chance to share context, catch bugs, and enforce conventions. It is also slow. A human reviewer needs time to load context, read the diff, and write feedback. An AI agent can do a first pass in seconds.

We tested OpenCode as a pre-review tool: before a human looked at a PR, the agent read the diff and produced a checklist of potential issues. The results were mixed but useful.

What the Agent Catches

The agent was consistently good at spotting:

• Missing error handling. Functions that returned values without checking for failure modes.
• Obscure variable names. Names like data or result that did not describe the contents.
• Dead code. Imports and variables that were no longer used.
• Test gaps. Code paths that had no corresponding test coverage.
• Style drift. Patterns that did not match the rest of the codebase.

These are exactly the kinds of comments that consume human review time but do not require deep product context.

What the Agent Misses

The agent struggled with anything requiring judgment outside the diff:

• Architectural fit. Whether a change aligned with the long-term direction of the system.
• Product intent. Whether the implementation matched the actual user need.
• Trade-off reasoning. When a simple but slightly slower solution was the right call for maintainability.
• Security context. Whether a pattern was safe in this specific domain.

Those remain human responsibilities. The agent is a linter-plus, not a senior engineer.

Knowledge Sharing as a Side Effect

One unexpected benefit was explanation. When asked, “summarize what this PR changes and why,” the agent produced a concise paragraph that was useful for:

• Standup updates
• Release notes
• Onboarding documents
• Handoffs between teams

The explanation required editing, but it was a better starting point than a blank page. For teams that struggle with documentation, this alone can justify the tool.

Practical Rollout

Start with one repository and one reviewer. Have the agent produce a comment draft for each PR, and let the human reviewer decide which comments to keep. After a few weeks, you will know which categories of feedback are reliable and which are noise.

Teams that already have strong conventions and fast tests get the most value. Teams without clear standards will find the agent’s feedback too generic to be useful.