Last week we wrote about how Tasks transformed Claude Code from a prompt assistant into a durable workflow engine.
That shift was about state. This week’s update treats the Agent Teams and approaches parallelism with Opus 4.6.
Claude Code now allows you to spawn multiple agents that work together inside the same project. On the surface, that sounds like a UX feature. Under the hood, it introduces distributed systems dynamics into your AI tooling.
If you’re building serious AI-assisted development workflows, this changes how you should think about architecture.
Let’s break it down.
From Single-Agent Loops to Concurrent Execution
Before Agent Teams, even with Tasks, you had one reasoning context:
User → Claude → Tool → Claude → Tool → Claude
Everything was sequential. Even if tasks were structured, execution was linear.
With Agent Teams enabled, the model shifts toward:
┌────────────┐
│ Lead Agent│
└──────┬─────┘
│
┌────────────┼────────────┐
│ │ │
┌────▼────┐ ┌─────▼────┐ ┌─────▼────┐
│Agent A │ │Agent B │ │Agent C │
└─────────┘ └──────────┘ └──────────┘
Each agent:
- Has its own context window
- Runs its own reasoning loop
- Executes its own tool calls
- Shares the same working directory
This isn’t role-play inside a single prompt, they’re independent reasoning processes.
How You Actually Enable Agent Teams
In Claude Code, Agent Teams are currently available as a research preview.
You enable them via environment configuration:
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
claude
Then you can orchestrate teammates through natural language:
Create an agent team to review this codebase.
Spawn 3 teammates:
1. security-reviewer – focus on src/auth/
2. perf-reviewer – analyze database queries
3. test-reviewer – assess test coverage
Have them share findings and challenge each other.
Claude handles spawning independent sessions behind the scenes.
What matters is what happens next.
The Architectural Reality: You Just Introduced Concurrency
Parallel agents introduce the same constraints you deal with in backend systems:
- Coordination
- Shared state
- Race conditions
- Resource amplification
If you treat Agent Teams like “more intelligence,” you’ll waste tokens. If you treat them like “concurrent workers,” you’ll start designing properly.
