Claude Opus 4.6 and Agent Teams

The Flagship Model Release

On February 5, 2026, Anthropic officially launched Claude Opus 4.6 (claude-opus-4-6), the most intelligent model in the Claude family to date. This flagship model was purpose-built for building agents and coding tasks, representing a significant leap forward in AI-assisted development capabilities.

Claude Opus 4.6 improves on its predecessor across multiple dimensions: it plans more carefully, sustains agentic tasks for longer durations, operates more reliably in large codebases, and has stronger code review and debugging skills to catch its own mistakes. The model supports a 200K context window by default, with a 1M token context window available in beta, up to 128K output tokens, extended thinking, and all existing Claude API features.

Agent Teams: Multi-Agent Collaboration

The centerpiece of this release is the Agent Teams feature, a research preview that fundamentally changes how developers can leverage Claude Code. Instead of having one AI handle everything serially from start to finish, developers can now spin up multiple Claude instances collaborating like a real development team.

Each sub-agent runs in its own tmux pane, handling a different part of the build while the orchestrator coordinates the overall task. This parallel execution model dramatically accelerates complex development workflows and mirrors how human development teams distribute work.

Anthropic demonstrated the power of this approach through an impressive experiment: 16 parallel Claude Agents wrote a 100,000-line C compiler (implemented in Rust) in just two weeks. The resulting compiler is capable of compiling the Linux 6.9 kernel with a 99% pass rate on the GCC test suite—a remarkable achievement that showcases the potential of multi-agent AI collaboration.

Persistent Memory System

Alongside the model release, Claude Code now automatically records and recalls memories as it works. Each project gets its own memory directory where Claude writes notes for itself based on what it discovers during sessions. This persistent memory system allows Claude to maintain context about project patterns, key commands, and user preferences across sessions.

The memory system includes frontmatter field support for agents with persistent memory scope (user, project, or local), enabling more sophisticated context retention. Additionally, new hook events were introduced specifically for multi-agent workflows: TeammateIdle and TaskCompleted allow better coordination between agents.

Skills and Extensibility

Skills located in .claude/skills/ within additional directories are now loaded automatically, improving discoverability and making it easier to extend Claude Code's capabilities. Plugin names are now included in skill descriptions, making it clearer which skills come from which sources.

Support for restricting which sub-agents can be spawned via Task(agent_type) syntax gives developers fine-grained control over multi-agent workflows, ensuring security and proper resource allocation.