Async agents often get mischaracterized. They’re not inherently asynchronous; their behavior depends on the user's actions. For instance, an agent can run tasks while the user does something else, making it seem async. However, if the user waits for a response, the interaction becomes synchronous. The article emphasizes that the distinction lies not in the agent itself, but in how it's used. The real potential of async agents emerges when they work concurrently. Users can delegate multiple tasks to different agents, but this leads to practical challenges. Agents need isolated environments to prevent overlap and conflicts within the same codebase. Solutions like git worktrees and isolated VMs are necessary to maintain this separation, allowing agents to function without interfering with each other. A deeper distinction is proposed between an async function and the runtime managing it. While every agent acts like an async function, a true async agent would be one that manages its own subagents and their tasks. This setup would allow users to focus on their intent rather than micromanaging multiple agents. The concept is still evolving, with early examples like Claude's Agent Teams and Gastown showcasing the potential but also highlighting current limitations in context sharing and inter-agent communication. As technology progresses, this model could become the standard in agent architecture.