makedream
8:43AI LABS
Log in to leave a comment
No posts yet
A single AI model will inevitably crumble when faced with large-scale projects. When trying to find bugs in a codebase with thousands of intertwined files or refactoring architecture, an AI working alone often loses context, falls into infinite loops, and ends up wasting tokens. This is the limitation of an isolated structure.
Anthropic's Claude 4.6 Opus-based Agent-Teams targets exactly this point. This isn't just about opening multiple chat windows. It is the realization of a software engineering "Swarm" where agents converse in real-time, share a local file system, and collaborate in parallel—just like a human development team. We reveal the practical strategies senior developers and tech leads should use to design and control this powerful team.
The decisive difference between Agent-Teams and traditional simple sub-agent methods lies in Shared State. While the past used a vertical structure where a main AI gave orders and merely received results, horizontal collaboration is now possible, with agents working together in the same office (local directory).
When a session begins, a .claude folder is created at the project root. This is the team's war room.
The most innovative feature is direct communication between agents. Using the SendMessage tool, a message is recorded in the receiving agent's inbox and injected as a system message in the next inference turn. In particular, the Heartbeat signals sent by team members at 2–4 second intervals allow the team leader to monitor member availability in real-time.
Simply commanding them to "collaborate" is irresponsible. Performance comes from assigning a clear persona and specific permissions to each agent. To find vulnerabilities in complex systems, design your team as follows:
| Role Name | Core Mission & Tools | Expected Deliverable |
|---|---|---|
| Security Auditor | Vulnerability scanning, static analysis | PoC code and reports |
| Performance Analyst | Memory profiling, query analysis | Bottleneck optimization proposals |
| Code Archeologist | History analysis, dependency mapping | Architecture diagrams |
| Test Engineer | Edge case test execution | Regression test reports |
To reduce interference between agents, you should utilize a Task Document structured like the one below as a prompt.
[Task Document Template]
/src/auth/* (Modification of files outside this range is prohibited)npm test and log the pass/fail status.Claude 4.6 Opus is powerful, but costs will soar if not managed. According to research data, utilizing Prompt Caching appropriately can reduce input token costs by up to 90%. There is no need to deploy the expensive Opus for every single task. Mix models based on the nature of the work.
Claude 4.6 Opus (Max Effort) - Overall architecture design and decision-making.Claude 4.6 Sonnet - The optimal balance of speed and accuracy.Claude 4.6 Haiku - Low-cost processing of large volumes of text.If the main agent tries to handle every trivial detail directly, switch the leader to coordination-only mode via Delegate Mode (Shift+Tab). Explicit constraints—telling the leader not to intervene until a team member's task is finished—are key to preventing token waste.
These are the environment setup steps to use this feature.
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1tmux environment is ideal for visually monitoring the agents' parallel work. Running with the claude --teammate-mode tmux option allows you to observe each team member's workspace in individual panels in real-time.Claude 4.6 Opus's 1 million token context window means agents can converse while holding tens of thousands of lines of legacy code and external documentation entirely in their "heads." In fact, in the MRCR v2 benchmark, this model's ability to find information within vast data improved by 76% compared to the previous generation.
A senior developer's core competitiveness is now shifting from the ability to write code directly to the ability to design and orchestrate AI agent teams. You must become an architect who sets the boundaries of complex systems and optimizes communication protocols between agents. Start with a small team of three or fewer and measure the productivity metrics for yourself.