Agentic Coding

Codex context management involves strategically organizing repository files, instructions, tools, and task details for AI coding agents. The most durable context belongs in version-controlled AGENTS.md files and reusable Skills, while task-specific context belongs in prompts. AGENTS.md serves as the foundation, discovered hierarchically from global to directory-specific levels, and should include project overview, coding standards, architecture, and constraints. Skills package repeatable workflows in directories with SKILL.md files and optional scripts, useful for multi-step processes like API endpoint creation or database migration. The Codex CLI supports Model Context Protocol (MCP) for connecting external tools like databases and browsers, but should be used only when agents need information outside the repository. Advanced patterns include context layering across multiple AGENTS.md levels, multi-agent parallel work using Git worktrees, and exploration-first prompts for complex tasks. Common mistakes include skipping durable project guidance, writing overly broad tasks, ignoring review steps, and adding unnecessary complexity with MCP when pure code-level context suffices.

ECC (Engineering Coordinated Cargo) is an open-source Claude Code plugin and orchestration system designed for agentic coding workflows. It provides 68 specialized agents, 292 reusable skills, and 94 command shims that manage context efficiently, enforce testing practices, and automate code review through a coordinated planning-test-implement-review-verify-remember-improve cycle. For solo SaaS founders using Claude Code and Cursor, ECC offers advanced context management through skills-first workflows, persistent memory across sessions, and specialized agents for planning, TDD, security review, and build error resolution. The system supports PRD-driven development with optional AGENTS.md and Skills frameworks that keep context focused and repeatable. Key features include continuous learning that extracts patterns into reusable workflows, AgentShield security scanning for agent configurations, and a unified memory vault for handoffs between harnesses. Installation uses guided setup, works with Claude Code's native plugin system, and provides language-specific rules and hooks that enforce standards without cluttering prompts. For developers seeking advanced vibe coding practices, ECC coordinates the engineer-agent partnership through deterministic verification loops and fresh-context review agents rather than relying on repeated reminders in chat.

Addy Osmani announced his upcoming O'Reilly book "Agentic Engineering," with chapters being released progressively starting now, with the full edition arriving in early 2027. Over the past six months, Osmani has been writing about agentic engineering fundamentals including loops, goals, context engineering, harnesses, verification, specs, autonomy, and software factories. He emphasizes that the book provides a framework for understanding how software engineering is evolving as AI agents take on increasingly larger pieces of work, covering practical aspects like delegating tasks, building harnesses, managing context limits, and determining what can safely be automated versus what requires human judgment. Osmani notes his views have evolved significantly through hands-on experimentation with bleeding-edge tools and workflows, with the expectation that many currently experimental practices will become standard by 2027.