A review that comments on everything gets muted
A good pull request review leaves three comments, not thirty. If an automated reviewer points out every deviation, the team skips them, and the important comment is lost among the others.
To decide which cases deserve a comment, you need the coding conventions your team enforces, those it is willing to tolerate, and the parts of the codebase where a small change can still matter. Context matters too. A missing test for an experimental endpoint and one for the billing path are the same kind of issue, with very different weight.
The agent leaves comments that state a point and then stop. It does not approve, merge, or block reviews. A human who has already been named runs the review.
How the agent is built in Agent Studio
Prompt skills state the criteria your team applies in reviews: the rules you enforce, not a style guide, and the conventions you have deliberately decided not to police.
A second prompt skill highlights the sensitive areas of the codebase, such as billing, authentication, and data migration. Weight varies with where a change is made because of those areas.
The system prompt caps the number of comments and requires them to be ranked. The agent has to choose what matters rather than produce a full list.
A comment must state the consequence, not only the deviation. If it only says that a convention might have been missed, it is noise. If it says that this path can now be used twice, it is a review.
A Webhook trigger receives pull request events. The GitHub node passes the diff, the related issue, and the recent file history to the Agent Chat node.
The GitHub node posts the comments back as review comments assigned to the agent, so they are not treated as coming from a colleague.
The agent neither merges nor approves any changes, and it blocks none. When a comment is posted, any item that involves a sensitive area is sent to a designated reviewer.
What this agent is built from
- Prompt skills: what your team enforces and what it has decided not to police.
- Prompt skills: the sensitive areas where a small change carries weight.
- System prompt: a capped, ranked set of comments rather than exhaustive output.
- System prompt: every comment states the consequence, not the deviation.
- Webhook (Triggers) with GitHub (Integration): the diff, the linked issue, and recent file history.
- Agent Chat (Util): comments posted attributed to the agent. No merge, no approval, no block.
- When an ActionFlow is enough: If you only need to check that tests, a linked issue, and a description exist, build the workflow. That check removes the most common review friction.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.