AI Pulse · Under the Radar

Most AI agent projects get tested in the weakest possible way. Did the workflow run without crashing? Did it return an answer? That tells you very little about whether the agent followed policy, chose the right action, used the right data, or gave the customer a useful result.

An indie developer has built Vizra ADK, a Laravel-native framework for building agents in PHP, alongside a vizra/evals package for testing them. The evaluation package uses Pest 5 and LLM-as-judge grading to assess whether an agent actually did the job it was meant to do, not simply whether the code executed. That matters because agent failures are often subtle. An agent can produce a polished answer that is wrong, take an action it should not have taken, or skip an important step without throwing an error.

For Laravel and PHP teams, this fills a real gap. Most agent tooling and evaluation examples are aimed at Python or JavaScript shops. Vizra has a modest star count, so this is early-stage software rather than a proven standard. Still, the niche is underserved and there appears to be very little direct competition.

If you run a business with internal AI assistants or customer-facing automation, the practical lesson is simple. Treat agent quality as something you test against real scenarios before deployment and after every change. This is the kind of thing we build into an AI command centre, where agents are monitored against the outcomes and rules that matter to the business.