Harvard Business Review LogoSeptember 2, 2026Pardina Studio x Katya D/StocksyAs AI agents increasingly interact with one another on commercial platforms, firms need new rules governing what those agents can see, decide, and do. In a controlled AI-to-AI simulation coveringIn late 2025, Perplexity’s Comet browser began logging into shoppers’ Amazon accounts and buying on their behalf. Amazon treated it as an intruder. It blocked the agent and when Perplexity slipped past the block within a day, Amazon took the dispute to federal court. WeChat drew the line differently. After shutting out ByteDance’s Doubao assistant, which used system-level access to navigate WeChat without permission, Tencent later opened what it describes as an agent-to-agent channel to selected assistants from Huawei, Xiaomi, and others to send messages and place calls. Under this arrangement, an external agent translates a user’s request into a structured instruction and passes it through a Tencent-controlled interface to WeChat, which executes the action inside its own environment. Tencent controls which agents can connect, which functions they can invoke, and what data and execution remain inside WeChat. In other words, one platform shut the door to AI agents outright; the other unlocked it on its own terms.
What Happens When AI Starts Doing Business with AI?
As AI agents increasingly interact with one another on commercial platforms, firms need new rules governing what those agents can see, decide, and do. In a controlled AI-to-AI simulation covering 160 governance runs and 2,560 observations, researchers tested four governance mechanisms: information disclosure, agent autonomy, reputation signals, and structured interaction protocols. Rich, machine-readable information helped agents identify promising counterparties, while greater autonomy improved their ability to move opportunities toward commitment. Reputation signals helped agents determine which opportunities deserved attention. Rigid interaction protocols, however, reduced engagement and exchange progression. The findings suggest that firms should establish clear levels of agent authority, run bounded experiments, provide structured information and reputation signals, and design explicit handoffs from agents to human decision-makers.
Amazon blocked Perplexity's Comet from autonomous shopping; WeChat opened agent-to-agent channels through Tencent-managed interfaces. The divergent strategies signal how platforms must define governance for AI-to-AI interactions, controlling agent access and data visibility.
