Agentic Coding

This Coursera course teaches building and debugging full-stack websites with Claude Code and mastering Claude Cowork agents for workflow automation across finance, legal, marketing, and data analysis. The curriculum covers transitioning from basic prompt engineering to orchestrating multi-agent AI systems, with hands-on demos connecting Claude to external tools like Gmail, Slack, and Salesforce via Model Context Protocol. The course emphasizes terminal-based autonomous orchestration, Agent Skills, and specialized plugins for data cleaning, dashboard creation, and financial modeling, enabling deployment of autonomous AI agent teams for end-to-end project management with live previews and automated reporting.

This is a comprehensive technical guide for Claude Code, the AI coding assistant. For the user's specific interests in agentic coding, advanced workflows, and best practices, here are the key relevant sections: The guide covers Agent Harness Engineering, which teaches how to identify controls required around agent loops and build reliable systems. It includes Loop & Graph Engineering for choosing between bounded feedback loops or explicit workflow graphs, and Context Engineering focusing on context budget, modular instructions, and assembly—critical for working with PRD documents and efficient AGENTS.md files as the user mentioned. The guide provides multiple advanced workflows including TDD with Claude, Spec-First Development, Plan-Driven Development, Best-of-N candidate comparison, and Agent Teams for coordinating multiple agents. For security and scaling, it covers Security Hardening, Sandbox Isolation, Production Safety, and Observability. The guide also offers reusable examples, decision frameworks (like MCP or CLI choice guides), and a structured approach to measuring context quality and task outcomes. It's available as an MCP server for integration into Claude Code, Cursor, and other coding assistants, making it directly applicable to daily developer productivity workflows.

Claude Code routines are saved prompt configurations that run on Anthropic's infrastructure and can be triggered by schedules, GitHub events (like pull requests), or API calls. The article demonstrates a practical workflow where a routine automatically generates README documentation when pull requests are created, showing how routines enable agentic coding automation without requiring local infrastructure. Key features include three trigger types (schedule, GitHub events, and API), notification options (push, email, Slack), and the ability to append additional instructions to routine executions via API requests. This approach allows developers to integrate AI-driven automation into their development lifecycle for tasks like code validation, documentation generation, and security scanning directly from GitHub or external orchestrators.

Agentic Awesome Skills (AAS) is a repository and control plane for managing reusable agent skills—sets of instructions that AI coding assistants like Claude Code and Cursor use to handle specialized tasks. The project's latest version (V17.8.4) introduces AAS Core, an agent-first preview system where Claude or Codex inspect your project, search a complete local catalog of 2,406+ skills, and choose exact skill IDs based on your needs. The system validates selections without ranking or recommending, records chosen stacks in reproducible aas-stack.json manifests, and allows immutable plan preview before any changes are applied. For solo SaaS developers working with Claude Code and Cursor, AAS offers structured workflows around skills (reusable SKILL.md playbooks organized by domain), bundles (curated role-based skill recommendations), and specialized plugins (focused marketplace distributions for web apps, security, data, DevOps, QA, and other domains). The recommended flow involves asking your AI agent to inspect your project with the local read-only AAS Core MCP configured, selecting exact skills from the full catalog, reviewing the stack manifest, validating via CLI, and previewing a deterministic plan in the Workbench before installation. Installation remains optional and approval-gated; the core value is reproducible, inspectable desired state rather than one-shot automatic setup.