Agent Team Work Zone: An Automated, Persistent Workspace for Long-Lived Coding Agent Teams
This paper proposes ATWZ (Agent Team Work Zone), a filesystem-based layer for Claude Code that treats agents as human employees by persisting their working states, skills, and inter-agent communications in dedicated files, thereby enabling state recovery after compaction, seamless team restoration, and reduced prompt-writing overhead.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Digital Memory Problem
Imagine you have a brilliant team of robot assistants, each an expert in a different field like coding, design, or debugging. You give them a massive project, and they get to work, chatting back and forth, making decisions, and building something amazing. But here's the catch: these robots only exist inside a temporary "bubble" of conversation. If you close the chat window, the internet cuts out, or the conversation gets too long and the robots have to summarize their history to make room for new words, the team effectively vanishes. Their progress, their specific roles, and the little notes they took on sticky pads are gone forever.
This is the world of modern AI coding agents. They are incredibly smart, but they suffer from a severe case of short-term memory loss. They can't remember who they were working with once the session ends, and they can't pick up exactly where they left off without a human stepping in to explain everything all over again. It's like hiring a construction crew, sending them home for the night, and then having to re-introduce them to the blueprints and the half-built house the next morning because they forgot everything. This paper tackles that exact frustration, proposing a way to give these digital teams a permanent "home base" so they can survive the night and wake up ready to work.
The Paper: Giving AI Teams a Permanent Home
The paper introduces a system called Agent Team Work Zone (ATWZ). Think of it as a magical filing cabinet and a set of strict house rules for your AI team. The core idea is simple but powerful: instead of letting the robots keep their important notes inside their fleeting "thought bubbles" (which disappear when the chat closes), the system forces them to write everything down in physical files on your computer's hard drive.
In this system, every AI agent gets its own personal "workstation"—a specific folder on your computer. Inside this folder, the agent keeps its role description, its current to-do list, the decisions it made, and a journal of what it just did. The paper argues that by treating these files like the agent's actual brain and memory, you can shut down the computer, close the chat, and even let the conversation get "compacted" (summarized to save space), and the team will still be safe.
The magic happens when you want to start working again. Instead of manually re-creating the team and re-explaining the project, you just run a single command: /reactivate-team. The system looks at the files in the workstations, reads the notes the agents left behind, and instantly rebuilds the team, handing each robot its specific job and the exact context it needs to continue. It's like pressing a "Resume" button that doesn't just bring back the main manager, but also reassembles the entire crew with their tools and blueprints perfectly organized.
The authors tested this by using the system to build itself—a process called "dogfooding." They found that while the AI agents themselves are still temporary and disappear when the chat ends, the files they leave behind are durable. This solves four major headaches:
- Lost Teams: You don't lose the team when the session closes; you just rebuild them from the files.
- Fuzzy Details: When a long conversation gets summarized (compacted), the agents don't lose their specific tasks because the details are saved in the files, not just in the chat history.
- Technical Debt: You don't get buried in old, unreadable chat logs. The decisions and progress are stored in clean, organized files that are easy to review.
- Prompt Writing: You don't have to rewrite long instructions every time you hand off a task. The agents just read the files the previous agent left for them.
The paper is careful to note that this isn't a magic fix that makes the AI agents themselves immortal. The agents are still "ephemeral," meaning they are just temporary processes. The innovation is the file-based layer that sits underneath them, acting as a persistent memory bank. The authors suggest that this approach turns a fragile, one-time conversation into a long-term, manageable project where the human is always in the loop, but the AI team can work across days and weeks without losing its mind.
It's important to know that this is a design manual, not a report on a massive scientific study with thousands of users. The authors haven't run big experiments to prove it's the "best" system in the world yet; they have built it, used it to build itself, and documented how it works and where it might break. They admit there are failure modes—like if an agent forgets to save its file, or if the system gets confused about which agent is which—but they have built safety nets (like automatic checkpoints) to catch these mistakes. The paper suggests that this file-first approach is a practical way to make AI teams useful for long-term projects, turning a fleeting chat into a lasting workforce.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.