Agent Mailbox Light

This Skill addresses the need for temporary, low-noise coordination between multiple AI agent workspaces, without permanently storing information in long-term memory. In practical development, developers often need different agents (e.g., responsible for frontend, backend, testing) to share temporary context prompts or status updates—such as notifying that an API is temporarily unavailable or reminding another agent that a task is completed. Traditional methods like chat or memory sharing are either too formal and token-heavy, or fail to persist across sessions, leading to inefficient coordination.

Usage is simple: just enable this Skill in an agent's workspace, and it will create a local .agent-mailbox directory containing three folders: inbox, keep, and archive. The sending agent scans other workspace directories and writes emails as Markdown files into the target workspace's inbox folder. The receiving agent automatically checks its own inbox upon startup or at task entry points, reads the latest few emails, extracts key information, and can then move useful emails to the keep folder for reference in subsequent sessions or archive them to the archive folder. The entire process requires no central server or network connection.

It is ideal for teams or individuals who need to coordinate multiple parallel AI agents to complete complex tasks—especially developers who have already created independent workspaces for different responsibilities (e.g., code development, documentation writing, testing) and want to establish a simple, non-intrusive communication mechanism between them. Project managers or full-stack developers can use it to pass temporary project status, configuration changes, or error warnings between agents with different functions.

It is recommended to uniformly integrate this mailbox-checking functionality when starting a project or configuring baseline workflows for each agent. Note that it is primarily designed as a "best-effort" broadcast, with no delivery guarantees or push notifications, making it unsuitable for conveying mandatory commands or critical task scheduling. Email content should be treated as "advisory prompts" for the local agent to evaluate; the agent should autonomously decide how to handle this information and strictly control the number of emails read each time to optimize token consumption.