My Agents Don't Get to Vibe 🤖

In the world of coding, tech enthusiasts often rely on multiple agents to streamline their workflows. However, author of the story warns against getting too attached to a single agent, as models and robots constantly evolve. The author's setup is intentionally simple and not tied to any specific agent. Skills define the workflow, and hooks ensure adherence to these skills.

Rather than giving an agent a single task like "Build feature X," the author prefers a council skill, where multiple agents independently analyze the problem, challenge the architecture, and turn the result into a technical issue. This includes elements such as SQL schemas, migrations, structs, types, classes, APIs, files, edge cases, and acceptance criteria.

The agent's job is to argue and alter the architecture while it's still cheap to do so. Once the issue is well-defined, the author hands it over to a builder, typically Grok Build. After implementation, there's an autoreview process, which includes review, findings, fixing the issues, and another round of review. This is far more effective than having the same robot write the code, review itself, congratulate itself, and ship.

The author has implemented this system across various projects, including acpbot.app, gegenlesen.app, and barkvisor.dev. The key difference lies in the verification of the final product. Instead of asking agents to simply say "Tests pass," the author enforces proof before pushing the code. Each project has a verification skill that knows how to confirm the code's functionality through running, clicking, querying, taking screenshots, and providing evidence.

If there's no proof, the code cannot be pushed. The author maintains a strict preflight check before pushing or opening a pull request (PR). While the robot can be creative during the process, the final step involves verifying that everything is in order before pushing the code. The author's workflow revolves around the agent, aiming to ensure its reliability.

The focus is on making the agent a reliable tool rather than striving to create the ultimate mega-agent. The author acknowledges the influence of Lauren Tan, Peter Steinberger, and Matt Pocock, whose work on skills, agent workflows, autoreview, and planning, specs, and engineering skills have shaped their approach.