Once multiple web AIs are placed in the same group chat, what really determines whether automation can continue is a control protocol connecting local agents to a browser extension. The solution demonstrated in the article involves at least four layers: a local daemon, a Chrome extension, and multiple web AIs. We believe this comes closer to a usable tool than simply “letting AIs chat with one another,” although its stability and maintenance costs still need to be validated.

What this is

This architecture addresses two core questions: who has the authority to make whom perform a task, and how the system recovers from failure. A local agent—a program that can read instruction files, execute commands, and advance tasks—sends commands over HTTP to a daemon accessible only from localhost. The daemon then passes those commands to a Chrome extension via WebSocket, and the extension operates the chat interfaces on the relevant websites.

Every request carries a unique ID, which the daemon uses to match execution results and timeout information. If the connection drops, the extension reconnects using exponential backoff, starting at 2 seconds and increasing to as much as 60 seconds. The design appears unsophisticated, but it tackles two practical constraints: Manifest V3 extensions cannot listen for connections as port-based servers, and the system may reclaim them after roughly 30 seconds of inactivity.

Industry view

We believe the article offers more than another method for orchestrating multiple models. It shifts the center of gravity in Agent products from model invocation to the “control plane”: permission boundaries, task confirmation, failure recovery, and observability determine whether automation can be trusted. Enterprises adopting this approach would still need to add identity authentication, operation auditing, and human-takeover mechanisms. A request being sent successfully does not mean the task on the web page has been completed correctly.

The risks are equally clear. Reliance on web interfaces means that platform redesigns, CAPTCHAs, expired logins, or risk-control prompts can interrupt a workflow. Localhost ports and shared tokens reduce the attack surface to some extent, but the coexistence of local agents, browser permissions, and web accounts still expands the credential-exposure surface. Moreover, the article mainly presents the design logic and code snippets; it does not yet provide measured results for cross-platform success rates, long-running stability, or maintenance costs.

Impact on regular people

For enterprise IT: This type of local control protocol may offer a low-cost way to test multi-model collaboration, but production deployment requires unified identity management, log retention, and permission approvals. Agents should not be given direct access to core accounts.

For individual professionals: It has the potential to connect web-based Q&A, document organization, and similar tasks into one automated workflow. At this stage, however, it is more of a developer tool. Users still need to check the results and remain responsible for actions such as sending, deleting, or submitting information.

For the consumer market: If this model becomes mature, consumers may be able to access multiple subscription-based AI services through a single entry point. Yet the ability of websites to support stable automated calls may also prompt platforms to strengthen authorization and anti-automation controls.