Agentic Coding

AI coding agents like Claude Code and Cursor lack built-in knowledge of AEM development best practices, so Adobe provides four complementary components to address this gap. AGENTS.md is a project-specific markdown file generated at your project root that grounds AI tools with essential AEM Cloud Service Java-stack expertise for every session. Agent Skills are reusable instruction sets encoding multi-step workflows: ensure-agents-md bootstraps tailored AGENTS.md files, create-component scaffolds complete AEM components, dispatcher handles Apache HTTPD configuration, migration guides modernization to AEM as a Cloud Service, workflow covers workflow model design and debugging, and code-assessment detects and fixes best-practice violations. Both Claude Code and Cursor support installing these skills via marketplace or CLI, and the create-component skill automatically detects project structure from pom.xml to streamline component generation. Additionally, the AEM Quickstart MCP server exposes live runtime data from local AEM SDK instances to support troubleshooting, while the Dispatcher MCP server enables runtime validation of Dispatcher configuration.

Claude Cowork extends Claude Code's agentic architecture to knowledge work tasks beyond coding, allowing users to describe complex multi-step outcomes and have Claude execute them autonomously. Key features include cloud-based task execution that persists across devices, three permission modes (Manual, Auto, and Skip) to control when Claude requests approval before actions, global and folder-level instructions for consistent task parameters, scheduled recurring tasks, and plugin support for role-specific customization. The system incorporates safety checks in Auto mode where Claude reviews actions for threats like data exfiltration before execution, though users working with high-stakes tasks are advised to stay close and review Claude's work or use Manual mode for greater control.

GitHub Copilot has implemented five key efficiency improvements for agentic coding workflows. Rather than optimizing individual tool calls, the focus is on completing full tasks efficiently: preserving useful context while removing repetitive output, eliminating unused formatting like line-number prefixes (which reduced inference cost by ~5%), compressing prompts without losing intended behaviors, and delivering background task results directly without requiring extra retrieval steps. The core principle is optimizing for completed task outcomes rather than token minimization—shorter responses can paradoxically increase total costs if agents must rerun commands or reopen saved output to recover missing information. Changes are validated through offline benchmarks and online experiments across different Copilot products (CLI, code review, app) to ensure results hold across varied workflows.