Agentic Coding

The article introduces agent harness engineering as the evolution beyond prompt and context engineering, defining it as a model plus five components: personalisation (global CLAUDE.md files), context (project-specific rules), action (tool connectors like Notion and Linear), memory (mistake corrections saved as permanent rules), and delegation (specialist agents for complex workflows). For Claude Code users, the harness is built through a folder structure at ~/.claude/ containing CLAUDE.md files, memory folders, skills, commands, and agents, while the same architecture applies in Cowork through settings and project-level configurations. The article emphasizes that the core discipline—stopping to document and save every mistake as a permanent rule—compounds over time, eventually enabling the agent to run jobs end-to-end with minimal human intervention, and notes that the harness itself is now the competitive moat rather than the underlying model.

oh-my-openagent is an open-source orchestration framework for agentic coding that addresses the user's specific interests in advanced AI agent workflows and optimization. The project features discipline agents like Sisyphus (orchestrator), Hephaestus (autonomous deep worker), and Prometheus (strategic planner) that work in parallel, eliminating manual model juggling across Claude Code, Cursor, and other harnesses. Key best practices highlighted include using hash-anchored edits to eliminate stale-line errors, auto-generating hierarchical AGENTS.md files via /init-deep for token efficiency, and leveraging the ultrawork command for complete workflow automation. The framework integrates LSP, AST-Grep, and built-in MCPs (web search, GitHub search, official docs) while maintaining full compatibility with Claude Code hooks, commands, and skills, enabling the efficient PRD and AGENTS.md workflows the user seeks without vendor lock-in.