AI Pulse · AI Trends Pulse

Anthropic just shipped Agent Teams inside Claude Code, and it’s the clearest signal yet that multi-agent systems are moving from research curiosity to production tool. The feature, available on Opus 4.6, lets one orchestrator manage between 2 and 16 sub-agents, each with its own context and tools, passing messages through a shared mailbox.

The proof case is striking. Anthropic ran 16 parallel instances to build a 100,000-line Rust C compiler over two weeks, then used it to compile and run Doom. Total token cost was around $20,000, according to the official documentation. That’s a real project, not a toy demo, and it gives you a rough cost benchmark for complex builds.

The practical guidance matters more than the headline number. Internal testing suggests 3 to 5 teammates is the productive range. Past five, coordination overhead eats the gains. That tracks with what we see in human teams, and it means the sweet spot for most businesses is probably three agents: one to plan, one to execute, one to validate or test.

Why this matters now

You can already imagine the shape: a research agent pulling data, a drafting agent writing the report, a QA agent checking citations and tone. Or a code agent building a feature, a test agent running edge cases, a docs agent updating the README. The orchestrator handles handoffs, the mailbox keeps state clean, and you get parallel work without the agents stepping on each other.

This is exactly the kind of workflow we’re building into systems like the Omni Command Centre, where multiple agents handle different parts of a business process and a single interface keeps you in control. The difference between this and a single chatbot is the difference between one person juggling tasks and a small team dividing them up. Agent Teams makes that division explicit, repeatable, and something you can run on your own infrastructure.