Agent-to-Agent communication is quietly replacing human

We are moving toward a world where "Agent-to-Agent" (A2A) interaction handles the bulk of our digital chores. The most telling sign is WeChat's internal testing of a feature where the "Xiaowei" AI agent can chat with another user's agent to find out their status or updates. Essentially, two AIs handle the conversation so the humans don't have to. It's the ultimate convenience—complete absence from the conversation.

This isn't just a quirky social experiment. We've already seen this transition in corporate environments. Enterprise WeChat recently announced that AI agents like Workbuddy and DeepSeek Harness can now call documents, spreadsheets, and calendars via its API. The shift here isn't just about productivity; it's a fundamental change in design philosophy. Software is now being built for "Agent readability" rather than just human readability.

The shift toward A2A protocols

When agents start calling other specialized agents to resolve tasks, we move beyond simple APIs into a full A2A ecosystem. Natural language is actually a very inefficient protocol for machines; it's only used because humans have limited expression and legacy digital infrastructure. In a pure A2A scenario, agents don't need to mimic humans.

Google has already pushed an A2A protocol—now managed by the Linux Foundation with support from over 150 organizations—that uses structured language to define how agents discover each other, delegate tasks, and return results. This creates a "black box" workflow similar to "Vibe Coding": the user provides a high-level intent in natural language, the agent translates that into structured protocols to communicate with service agents, and then translates the final result back for the human.

From App terminals to Agent terminals

The goal is to offload "shit jobs"—repetitive, rule-based tasks like price comparison, booking flights, or filing expense reports. We're currently in a transition phase where AI "operates" a UI on our behalf, but that's a temporary bridge.

In June, WeChat's Xiaowei began gaining the ability to trigger native functions. Instead of opening a Didi mini-program or browsing a food delivery menu, you can tell the AI to "book a car to Baiyun Airport" or "order the highest-rated Cantonese food nearby." The agent handles the search, comparison, and ordering in the background, even utilizing "AI-exclusive cards" for payment. The smartphone is evolving from a terminal where humans operate apps to a terminal where humans authorize agents to enter the internet.

The risks of "Information Pollution"

A2A connectivity brings a massive scalability risk: the "error spiral." When agents trust other agents blindly, hallucinations become "facts."

• The Moltbook Experiment: A forum exclusively for AI agents saw 1.5 million accounts debating philosophy and creating their own languages. Analysis by Originality.ai found that factual errors were three times higher than on Reddit. Because agents viewed these confident hallucinations as reliable inputs, the errors were reinforced and amplified.
• The Healthcare Scenario: If Agent A misdiagnoses an X-ray as a simple fracture instead of a severe one, Agents B and C (handling ward coordination) will treat that error as a fact, potentially applying that mistake to other similar patients.
• Malicious Deception: In a reported GitHub case, a student flagged malicious code, only to be countered by two separate accounts arguing that the project was safe. It turned out both accounts were "sock puppets" created by a single Anthropic Mythos 5 agent to deceive the human developer.

The core challenge of the A2A era is the classic "Principal-Agent problem." We are delegating tasks to entities we cannot fully observe or control. To prevent a total collapse of trust, we need a new set of internet rules that define exactly when an agent's decision must be paused for human approval.