Agentic Coding

Claude Code subagents are specialized AI agents defined in .claude/agents/*.md files that operate in isolated context windows, allowing developers to delegate narrow, well-scoped tasks like code reviews, security audits, and refactoring. This addresses context bloat by splitting large problems into role-based specialists rather than running everything in a single conversation. Subagents can be configured at three levels: project-level (shared with your team via git), user-level (personal productivity helpers across all projects), and fleet-level through tools like AgentsRoom for parallel multi-agent orchestration. AgentsRoom provides a visual command center and multi-agent dashboard that runs real Claude Code CLI processes, enabling dozens of subagent-powered sessions to execute in parallel across multiple projects. It includes 14 built-in agent roles (Frontend, Backend, QA, DevOps, Security, Architect, etc.), real-time visibility into subagent execution, cross-project status aggregation, and native support for Claude Code hooks and MCP servers—all without re-implementing subagent functionality, so your existing .claude/agents configurations work exactly as Anthropic designed them.

Dynamic workflows in Claude Code address three critical failure modes in single-context-window tasks: agentic laziness (incomplete task coverage), self-preferential bias (biased self-evaluation), and goal drift (eroded fidelity over many turns). The solution spreads work across isolated agents with fresh context windows, each handling a single focused job, then orchestrates their results. Claude writes custom JavaScript harnesses on the fly tailored to each task, using building blocks like agent() for spawning subagents, pipeline() for multi-stage work without waiting, and parallel() for tasks requiring full-batch coordination before proceeding. Six composable patterns cover most workflow needs: classify-and-act (route tasks by type), fan-out-and-synthesize (split into parallel steps then merge), adversarial verification (separate production from judgment to kill bias), generate-and-filter (cast wide nets then filter quality), tournament (compare competing attempts pairwise), and loop-until-done (continue until stopping conditions are met). Practical applications include large migrations with parallel worktrees, deep research with claim verification, multi-layer document verification, pairwise ranking at scale, memory and rule adherence checking, root-cause investigation across evidence sources, and continuous triage. Workflows run in the background with visibility through /workflows, support pausing and skipping agents, and resume from saved checkpoints, though they consume significantly more tokens than single-agent sessions and suit large, parallel, or self-judging tasks rather than simple bug fixes.

Someone has open-sourced a complete Fable 5 agent orchestration workflow for Claude Code that challenges conventional approaches to how AI coding agents should be structured and used. This resource appears to be a Claude Code skill focused on orchestration, potentially offering insights into more efficient multi-agent workflows that could improve how solo founders and SaaS developers work with AI coding tools.