@anorth_chen: Peter's article clarifies the fundamental boundary between cloud agents and desktop agents: once the agent leaves the user's computer, the problem shifts from a framework issue to an infrastructure contract. Desktop agents come with many implicit assumptions: a trusted local filesystem, trusted keys in the environment, an online user, the ability to manually retry on failure, and the ability to reinstall the environment if it breaks.

Peter's article clearly explains the fundamental boundary between cloud agents and desktop agents: once the agent leaves the user's computer, the problem shifts from a framework issue to an infrastructure contract. Desktop agents come with many implicit assumptions: a trusted local filesystem, trusted keys in the environment, an online user, the ability to manually retry on failure, and the ability to reinstall the environment if it breaks. But for cloud agents, they need to run in unsupervised, shared hardware environments that may be affected by prompt injection, while also supporting being invoked by cron, APIs, and other agents. Therefore, from a technical paradigm perspective, agent runtime will increasingly resemble a small OS in the future. Many users and friends have asked us about the security management of secrets, and this article provides a great answer to our approach: We don't try to solve sandbox security, as that would trap us in a cat-and-mouse game with hackers and drag down our business. Instead, we take a different approach: assume the sandbox has already been compromised, and ensure that long-lived secrets never enter the sandbox execution boundary. This way, the payoff for hackers is not worth their time and effort, and they will naturally give up. Peter's technical article perfectly illustrates what I previously said about the difficulty of making agent products production-ready: during the demo phase, you only need to manage context and the number of tools; but in production, the reliability of the product depends on these boring things: snapshots, JWT, IP allowlists, and consistent execution pipelines for billing/logging/observability. An agent is a function with a natural language interface. The business capability of an agent depends on the user's know-how. As a platform, we help users solve infrastructure challenges such as trigger surfaces, runtime, and security boundaries.