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

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

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