An orchestrated multi-agent system that turns written test cases into executable ROBOT framework scripts — combining a reasoning model and a code generation model, then reviewing its own output before it ships.
Test automation teams were hand-writing ROBOT framework scripts from written test cases — slow, repetitive, and inconsistent between authors. Existing scripts that solved the same problem already existed in the project, but nobody could find them, so the same work was done repeatedly.
I built an orchestrated workflow that splits the job between a reasoning model and a code generation model, so the hard thinking and the syntax are not competing for the same context. A production RAG layer on QdrantDB with KNN vector search surfaces existing scripts before anything new is generated, which turned reuse from an aspiration into the default path. The agent framework itself is config-first and model-agnostic across OpenAI and Anthropic, so swapping or adding a model is a config change rather than a rewrite. Semantic Kernel handles the enterprise integration, with custom plugins for the domain-specific pieces.
Query relevancy for existing scripts improved enough that reuse became routine rather than accidental. The config-first design meaningfully shortened the path from a new model being available to it being deployed.