Phase 6 of the Jarvis AI Platform introduces agents, enabling multi-step reasoning and tool utilization beyond single requests. Traditional chatbots struggle with complex tasks requiring planning, but agents employ the ReACT pattern, a loop of thinking, acting, and observing. This new functionality was implemented as a separate orchestration layer, keeping the existing chat pipeline unchanged. A four-layer system, including an AgentController and AgentPlanner, manages agent execution and tool integration. The system uses structured prompts as contracts and robust parsing to handle AI's output precisely. Safety limits like maximum steps and timeouts prevent agents from running indefinitely. Design decisions focused on precise tool matching, streaming agent events, and graceful handling of client disconnects. Agent state transitions are managed by the domain model using compare-and-set updates to prevent race conditions. The REST API supports both streaming and asynchronous agent execution, allowing complex, multi-tool requests to be processed efficiently. Lessons learned emphasize structured prompting, graceful degradation, domain model enforcement, and compare-and-set for concurrency. Phase 7 will focus on building the web interface to integrate all previously developed features.