From a Review Skill to a Small .NET PR Review Agent
noteSeptember 30, 2026Updated October 7, 2026By Ali Haghighi
After defining the .NET Change Reviewer skill, the next problem was not how to make the model smarter. It was how to let the workflow gather enough context without turning a focused reviewer into a general repository automation system.
That gave me a useful distinction. The skill defines how to reason about a change. The orchestration layer decides what context to inspect, when uncertainty should become a question and which actions remain outside its authority.
I used that boundary to turn the small review skill into a first agent workflow.
The agent does less than the name suggests
The .NET PR Review Agent starts from a pull-request diff. It may inspect a small number of directly relevant files, applies the review skill, challenges unsupported findings and returns structured JSON.
It cannot post comments, modify code, approve a pull request or merge anything.
That limitation is intentional. Adding write actions before evaluating read-only behavior would make the experiment harder to reason about.
Context is part of the design
Giving an agent an entire repository is convenient, but it makes cost and behavior harder to understand. V1 starts with the diff and uses a small context budget for interfaces, tests and directly related implementation.
The number is a guardrail rather than a universal rule. When the available context is insufficient, the expected behavior is to identify what is missing rather than guess.
Questions are first-class output
A review finding says the supplied evidence supports a problem. A question says the answer depends on a contract or context the agent does not have.
Keeping those separate is useful for cases such as stream ownership. Disposing a stream can be correct or incorrect depending on the API contract.
Evaluation before autonomy
The initial scenarios cover dropped cancellation, ambiguous ownership, bounded streaming, a compatibility break, a clean change and missing context.
The next step is not to give the agent more permissions. It is to execute these scenarios repeatedly, record false positives and missed findings, and refine the workflow based on evidence.
The workflow and output contract are linked from the AI Lab.