Agentic AI Software Testing and Development with Dragan Spiridonov

The conversation centers on Dragan Spiridonov’s development of an open-source Agentic QE orchestration system that applies specialized agents and skills across the software development lifecycle. It traces how he moved from standard chat-based LLM use to prompt engineering, then context engineering, and finally agentic methods that produce over 90 percent acceptable outputs on quality tasks.

From Chat Interfaces to Agentic Execution

Dragan describes a clear sequence of capability gains. Early experiments with ChatGPT, Gemini, and Claude in chat windows yielded roughly 50-60 percent usable results because prompts alone could not maintain project context. Adding context through memory banks in tools such as Cursor and Roo Code raised reliability above 80 percent. The decisive step occurred when he adopted agentic engineering: agents first observe the environment, collect data, reason selectively by calling LLMs only at decision points, and then execute deterministic logic or tool calls toward an explicit goal. The human role shifts to reviewing outputs and intervening when patterns deviate.

This structure separates concerns. Deterministic code inside agents handles repeatable steps without LLM cost or variability, while LLM calls occur only at crossroads. The result is traceable behavior that can be observed in real time rather than opaque end-to-end generation.