Agentic Coding

Agentic engineering is evolving into two distinct generations. The first generation, currently dominant, uses deterministic workflows where engineers design fixed paths with specialized agents handling narrow tasks in sequence—like triage, diagnosis, remediation, and verification stages. The second generation moves toward autonomous agents that receive a goal and determine their own path, similar to how Claude Code operates by reading context, planning approaches, executing tool calls, and iterating until goals are met. This shift, termed "loop engineering," requires engineers to design the systems that prompt agents rather than manually prompting them. The foundation for this transition includes building a context lake for agent reasoning, an agent registry, a skills registry, an MCP hub for tool access, and governance layers with identity and blast radius limits. Teams adopting agentic platforms now for deterministic workflows are simultaneously building the infrastructure needed for autonomous 2.0 systems, positioning themselves to transition as models and trust mature.

AAS Core provides local, agent-owned skill stack management for Claude Code and Cursor. Version 15.5.1 introduces AAS Core for complete local catalog search where Claude or Codex inspects your project and selects exact skills from a local catalog without uploading code. The system validates selections in memory via a read-only compose_stack tool, then persists an aas-stack.json manifest and optional evidence file for human review before any changes. The workflow emphasizes agent-first composition with human control: agents evaluate full project surface across architecture, testing, security, and deployment, then search and compare skill candidates to cover each capability without redundancy. Validation and immutable planning are supported; apply and recovery remain experimental. For Claude Code and Cursor users, the MCP setup lets agents search the complete 1,992+ skill catalog, and the CLI or client persists the approved stack as a reproducible, inspectable artifact. Specialized plugins package proven skill sets for web, security, data, DevOps, and other domains; direct install, bundles, and workflows remain available for other distribution paths.

I rebuilt a $13M quit-drinking app using Claude Code without opening Xcode, demonstrating that the core structure is straightforward: a streak counter, milestone screens, and a paywall. This approach showcases practical AI-assisted development for building functional SaaS applications efficiently without traditional development environments.