Building 33 AI Agents on One Engine: A Look Inside Five of Them

Over the past year, the author built 33 AI agents on a single engine—a single Docker image and codebase, with no core modifications. The agents range from a chess coach for a child to assistants for a wife, tutors, and corporate bots for metallurgical and energy companies. The main variable is where each agent draws the line between what code calculates and what the model says. For personal agents, the model may compute directly, but in corporate ones it never outputs final numbers; it only sets input parameters, which poses a risk. Tool schemas start to dominate the prompt after the third dozen tools; for a CRM admin with 331 tools, JSON schemas amount to 302 KB or about 75,000 tokens in the prefix of each request. Three config tweaks reduced prompt tokens from 15,266 to 1,424 and median response from 17 to 7.7 seconds. The engine is Hermes Agent from NousResearch, with a young codebase from July 2025. The agents differ mainly in their persona and the set of tools; the article provides tables comparing prompt sizes and model choices. Some failures are also discussed, such as a balance discrepancy and 5,863 lines of Python never invoked in production.