Claude Code Agent Teams represent a significant evolution in how AI agents are designed and deployed, marking a shift from rigid, code-defined workflows to dynamic, conversation-driven team structures. Unlike traditional agent architectures that rely on predefined configurations, YAML files, or workflow schemas, Agent Teams operate through natural language prompts. You simply describe the team you want—its roles, responsibilities, and collaboration style—and Claude Code dynamically creates and manages the agents in real time. This approach eliminates the need for configuration files or complex orchestration logic. Instead, a single setting in settings.json enables the feature, and a conversational prompt defines the entire team. The model interprets the description, spawns independent Claude instances for each teammate, manages task assignment, and handles communication between them—all autonomously. One of the most striking aspects is that teammates don’t inherit the lead agent’s conversation history. Each starts fresh with project context loaded from CLAUDE.md, MCP servers, and available skills. They communicate directly, claim tasks using file locking to avoid conflicts, and unblock dependent tasks automatically when work is completed. This direct peer-to-peer interaction removes the bottleneck of having all communication route through a central supervisor, enabling more fluid, realistic collaboration. The model behind this—Opus 4.6—is specifically optimized for agentic behavior. It doesn’t just generate text; it parses prompts, decides which tools to use (like TaskCreate or mailbox messaging), manages coordination across multiple turns, and maintains state without manual intervention. The simplicity of markdown prompts is deceptive. Under the hood, it’s a sophisticated system where the model acts as both the orchestrator and the executor. There’s a clear tradeoff between ephemeral and durable agents. Ephemeral teams are ideal for quick, one-off tasks—reviewing a PR, debugging code, or exploring a codebase. They’re fast, disposable, and require no setup. Durable agents, on the other hand, are versioned, tested, and reusable—perfect for CI pipelines or production workflows. Agent Teams sits in the middle: interactive, exploratory, and lightweight. The real power lies in the spectrum of control. You can be prescriptive (“spawn security, performance, and test reviewers”), guided (“use multiple specialists, max four agents”), or open (“review this PR”). The level of detail depends on your need for predictability, cost control, and structure. This paradigm reflects a broader trend: models are moving up the stack. As AI becomes more capable, the need for external frameworks diminishes. The future may see agentic applications reduced to simple markdown files, with the model handling everything from team formation to execution. Anthropic’s bet is clear: as models grow smarter, the orchestration layer becomes redundant for most use cases. Agent Teams is not a replacement for production frameworks but a new way to experiment, iterate, and prototype at speed—without boilerplate. It’s a glimpse into a future where the model itself is the framework.