Anthropic released Opus 4.6 on February 5, introducing agent teams: multi-agent systems that split tasks across parallel agents for faster coordination. The feature is available in research preview for API users and subscribers.
What ships
Opus 4.6 adds three capabilities. Agent teams use orchestrator-worker patterns where lead agents delegate to subagents, handling planning, tool execution, and error recovery. The model now supports 1M token context windows, matching Sonnet 4 and 4.5, enabling larger codebases and documents. PowerPoint integration moves from file export to a native side panel.
According to Anthropic, agent teams coordinate in parallel rather than working sequentially. Subagents condense work to 1,000-2,000 tokens per summary, according to production testing.
The deployment reality
Constellation Research flags three production risks CIOs should watch. Agents misjudge effort without explicit scaling rules. Subagents need detailed instructions or fail silently. Uncoordinated updates risk outages. The firm recommends observability beyond conversation logs.
Internal testing data shows tool description improvements cut task completion time by 40%. The pattern is clear: poor tool descriptions cause agent failures. This isn't a prompt engineering problem, it's an architecture problem.
Enterprise context
Opus 4.6 shifts Claude from software-focused to broader knowledge work: product managers, financial analysts, business operations. Anthropic emphasizes evals (model and human graders) plus modular skills (instructions and scripts) for reliability. Safeguards include retries, checkpoints, and human oversight gates.
The comparison to existing frameworks matters. LangGraph, CrewAI, and Microsoft's Agent Framework handle similar orchestration patterns. OpenAI's Swarm offers lightweight coordination. The real question for enterprise teams: does Anthropic's integrated approach justify vendor lock-in versus open-source flexibility?
Rate limits and context windows create practical constraints. Claude's API has tiered rate limits and weekly token caps that affect agent team scaling. Production deployments need buffer strategies.
What to watch
Agent teams move multi-agent systems from research to production. The technology works. The operational patterns aren't settled yet.