Agentic Coding

The supervisor pattern — where a lead agent orchestrates specialist subagents — has emerged as the dominant multi-agent architecture in coding because development provides the fastest feedback loop for detecting architectural failures: code either compiles or it doesn't. Claude Code's design reflects this reality, with a lead agent spawning subagents through two primitives: forking (inheriting full context for investigations building on prior work) and spawning fresh (creating independent reviewers unbiased by the implementer's assumptions). A typical session involves the lead agent dispatching an Explore agent to map the codebase, synthesizing findings into a plan, then delegating implementation before a fresh review agent checks the result — a hierarchy with clear synthesis ownership that avoids the coordination failures of peer-agent swarms. Enterprise deployments at Klarna and Cisco are now adopting this same supervisor pattern for workflow automation and customer service, suggesting that the architecture proven under coding's harsh feedback loop generalizes well to slower domains. The key lesson for teams building multi-agent systems outside developer tools is to deliberately manufacture faster feedback loops through synthetic testing rather than relying on production failures, since the patterns that survive coding's immediate build-or-fail signal are worth trusting as defaults even in domains with delayed feedback.

Agent Skills is a production-grade framework that encodes 24 structured workflows for AI coding agents like Claude Code and Cursor. The pack maps the full development lifecycle — from defining requirements through shipping to production — using slash commands (/spec, /plan, /build, /test, /review, /ship) that automatically activate relevant skills. Each skill includes step-by-step processes, anti-rationalization tables that counter common shortcuts, and verification gates that enforce practices like test-driven development, spec-before-code, and code review. The framework embeds engineering best practices from Google's culture, including Hyrum's Law for API design, the Beyonce Rule for testing, and Chesterton's Fence for code simplification. Installation is tool-agnostic: use npx for any agent, or native integrations for Claude Code (via /plugin marketplace add), Cursor (.cursor/skills/), Gemini CLI, Windsurf, GitHub Copilot, and others. The /build command removes manual handoffs between planning and implementation — you approve the plan once, then it runs autonomously while still enforcing test-driven commits and pausing on failures. Specialized agent personas provide targeted reviews (code-reviewer, test-engineer, security-auditor, web-performance-auditor) and work with orchestration patterns that keep multi-agent workflows aligned and prevent personas from invoking other personas.

Agent Fleet is a multi-agent orchestration system for AI coding agents that operates specialist teams rather than single chat sessions. Built pi-first, it routes work through a thin dispatcher to specialist agents (planner, builder, test-engineer, code-reviewer, documenter) under a Verification Contract, with agents communicating over a shared messaging plane. Key features include tiled workspaces you can snapshot and resume, a library of 29 production-grade lifecycle skills and 15 personas that enforce discipline across planning, building, testing, reviewing, and shipping phases, and the ability to mix multiple model subscriptions (Claude, ChatGPT, GitHub Copilot, Ollama, local GPU models) within a single fleet. The dispatcher avoids context drowning by keeping research output and specialist chatter on disk behind file paths rather than in the orchestrator's context window. Claude Code appears as a first-class coms peer for cross-model review and answering other agents' questions. Control surfaces include herdr for tiled workspaces with snapshot/resume, a Hermes Desktop plugin showing live fleet state and blocking agents, and a Hermes relay that pipes questions to Telegram for remote human approval. For solo SaaS founders using Claude Code and Cursor daily, this enables efficient multi-agent workflows with advanced skills like spec-driven development, test-driven development, incremental implementation, code review, and production shipping practices embedded directly into agent behavior.

Amp is a coding agent and development environment featuring "orbs" (remote machines where agents run unsupervised). Each thread spawns a fresh orb with code, plugins, and tools—eliminating workspaces management and port conflicts. Agents continue working after you close your laptop, and you can spawn multiple agents for tasks ranging from small fixes to complete features, managing them via web, terminal, or mobile. Recent updates include support for multiple repositories, friendly URLs for sharing orbs, MCP server integration, and setup without initial commits.