Specialized agent harnesses enable non-technical founders to drive development
Bastian Buch describes a multi-agent system that allows non-technical founders to contribute to frontend development through natural language feedback. The approach uses specialized 'harnesses' to abstract complexity, avoiding the context overload associated with direct LLM prompting.
Bastian Buch describes a workflow built for a RegTech startup founder to allow her to contribute to frontend development without coding. The system uses Pen.dev for design system visibility and Linear as the primary interface. A design agent creates issues in Linear with design tokens and components, sending screenshots for review. The founder provides feedback in natural language, and the agent iterates on the design, sending new screenshots until approved. Once approved, the changes are automatically opened as a Pull Request in GitHub for review, bypassing the need for the founder to use complex coding tools like Claude Code.
Buch contrasts this with 'vibe coding' or direct LLM prompting, which he argues produces average results due to context overload or lack of structure. He advocates for 'harnesses,' specialized agent setups with distinct roles such as director, narrator, and continuity manager, that abstract complexity. He references Matt Pocock’s advice to separate code generation from code refinement using specialized agents, noting that this approach mirrors software engineering abstractions where goals are set for agents rather than step-by-step instructions.