Under the Hood: The Architecture Behind Claude Code Agents

Everyone is talking about Claude Code and agent-based code tools. But to really work well with them, you need to understand not just the agent itself - but the harness around it: the context layer, memory, tools, permissions, and rules that enable it to function reliably. In this session, Yaara Cohen will explain what a harness is through the example of Claude Code, describe the architecture surrounding code agents, and how to build a harness and what it includes. What we will unpack in the session: * What the central loop and agent management look like and how this differs from a regular chat with a model * Why context engineering is becoming a central part of the work, and not just 'writing a good prompt.' * How memory, permissions, and work rules affect the quality and safety of development. * What the role of subagents and work stages is when building complex systems. * Why the future of agent engineering lies less in prompts and more in harnesses. Who is this for: * Developers working with Claude Code, Cursor, or agent-based code tools who want to level up. * Founders and builders creating products using code tools who want to understand how to work with them systematically. * Product and technology professionals wanting to understand what the architecture of agent systems looks like beyond a demo. About the speaker: Yaara Cohen is an AI Developer at Stealth, currently working on an agent management platform. She has end-to-end experience in building SaaS products and technology projects, including previous experience at Wix.