NEW Claude’s Agent Swarm UPDATE!

On March 31st, 2026, Anthropic experienced a major accidental source code leak when an intern at a blockchain security firm discovered that Claude Code's npm registry contained a debugging file pointing to 5,000+ lines of TypeScript code. This represented the second major leak in a week, following a previous disclosure of 3,000 internal files revealing an unreleased model called Mythos. The leaked code exposed numerous feature-flagged capabilities that Anthropic had built but not yet released. The most significant discovery was a complete multi-agent orchestration system called 'teammate tool' with 13 distinct operations for spawning and coordinating AI agent teams. This system allows a lead agent to break down tasks while specialist agents work in parallel on different components like frontend, backend, testing, and documentation through shared taskboards and messaging. The leak also revealed 'Chyros,' a daemon mode background agent mentioned over 150 times in the source code, designed to watch files and continue working even when users are idle. Paired with this is 'AutoDream,' a memory consolidation process that organizes and learns during downtime, and 'Ultra Plan,' an enhanced planning mode for complex long-term tasks. The code exposed Anthropic's internal model roadmap, showing that 'Capiara' is the code name for Claude 4.6 variant, 'Fenic' maps to Opus 4.6, and 'Numbat' is still in pre-launch. Internal benchmarks revealed accuracy challenges, with the latest Capiara version showing a 29-30% false claims rate. Agent teams became officially available as an experimental feature in February 2026, though the infrastructure was built months earlier. The system addresses context limits by giving each agent independent context windows, with Anthropic's research showing token usage explains 80% of performance variance in complex tasks.