Post

We've been building our own agents to help us operationally and they all run on open weight models, mostly GLM The newest is a code review agent that runs GLM-5.3 Flash. The Prisma team is already favoring it over the one we pay for and the core was built in a couple afternoons

Tyler Benfield

@rtbenfield.dev

We use Mastra for its excellent primitives and telemetry. It gives us a good foundation, is model agnostic, and improves every week. The skill and docs are great, so we unleash our agents to build agents. Highly recommend

8/9/26, 4:50 pm0 reposts1 like

Tyler Benfield

@rtbenfield.dev

8 Sept

The code review agent starts from a GitHub webhook, grabs some context, clones into a sandbox, generates a git diff, then gives the sandbox to an agent to perform the review

Tyler Benfield

@rtbenfield.dev

8 Sept

The agent has tools to report feedback and mark files as complete. If it ends the review without all files marked complete, a guard makes it continue. This all happens in a Mastra workflow. A few more things for replying to comments and such, but that's the gist of it