Anthropic's Agent Teams release makes agentic workflows look less like “prompting” and more like…

Anthropic's Agent Teams release makes agentic workflows look less like “prompting” and more like management and org design.

The shift from “one assistant” to “a small org” is going to be a massive change in how we interact with, manage and delegate tasks to agents.

Teams fixes one of the big pain points of subagents where now, not only are the agents fully independent sessions with their own context windows, but they can message each other directly and interactively rather than just reporting an outcome to the lead.

Work ends up coordinated via a shared task list with dependencies + file locking to avoid stepping on each other's virtual toes and you can force the main agent into delegate mode so it only coordinates rather than trying to do everything itself.

I'm especially interested to see how the second order effects play out:

Anthropic suggests having agents argue to disprove each other - dissent for debugging!

Quality as an org pattern: separate builder / tester / security reviewer / architect roles aren’t just prompt role suggestions when they can challenge each other before merge. This gives a new kind of continuous review loop.

Tokens scale with active teammates, and task sizing matters so this definitely will melt some glaciers and isn’t a free gimme, it’s a tradeoff that people will have do decide how and when to use.

I've started testing with a simple setup: designer + builder + tester + security reviewer, all allowed to debate each other, with plan-approval gates for anything risky.

Anyone else tried Agent Teams yet?

What role split gave you the biggest jump: speed, quality, or fewer “looks right to me” bugs?