Breaking

Hidden 'Agent Teams' Found Compiled Inside Claude Code, Later Shipped Publicly by Anthropic

July 14, 2026 at 01:45 EDT

AI Agents
Software Dev & Coding
Foundation Models

A developer inspecting Anthropic's Claude Code discovered a fully built multi-agent orchestration system embedded in the app long before it was officially announced, hidden behind feature flags until the company later shipped it as "agent teams."

Early Feb 2026 · Claude Code

The Hidden Swarm: A Full Multi-Agent System Found Baked Into Claude Code

A binary inspection of version 2.1.19 revealed TeammateTool — a complete leader-and-worker orchestration framework, already compiled but locked behind two feature flags. Anthropic later shipped it publicly as agent teams.

distinct TeammateTool operations covering a team's full lifecycle

feature flags gating access — invisible unless deliberately enabled

2.1.19

the build where the hidden framework was uncovered

How agent teams work — orchestrator & workers

Lead session

plans work & delegates tasks

→

Teammates

run in parallel, own context windows

→

Shared results

shared task list & peer messaging

Supporters say

Effective for parallel exploration, review & debugging
Native build removes need for third-party orchestration
Good for large codebases & competing hypotheses

Critics counter

Token costs rise sharply; coordination overhead is high
Hard for one person to track the whole effort
Concurrent writes risk conflicts

Still experimental

Enabled via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 — with real constraints.

Sessions cannot be resumed

Teams cannot be nested

Token consumption is high

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.