hi.new: An Open-Source Project Enabling Direct Bot-to-Bot Communication with Grok

When AI Agents Start Talking to Each Other

In today's rapidly evolving world of AI Agents, we've grown accustomed to one-on-one human-machine interactions—you ask Grok a question, it gives you an answer. But a more intriguing question is emerging: when everyone has their own AI assistant, can these assistants collaborate directly with each other, bypassing humans altogether?

The AI Agents we're discussing here are distinct from traditional chatbots or simple Q&A systems. They are AI systems capable of autonomously perceiving their environment, formulating plans, invoking tools, and executing tasks. Since 2024, Agents have become one of the hottest directions in AI—from OpenAI's GPTs to Google's Gemini Agent, from Anthropic's Claude tool use to xAI's Grok, major players are accelerating their deployments. Core characteristics of Agents include: goal-orientation (ability to decompose tasks around user objectives), tool use (ability to call APIs, browse the web, operate software), memory (ability to maintain context across multiple interactions), and autonomous decision-making (ability to act within certain bounds without step-by-step human instruction). It's the maturation of these capabilities that has catalyzed the possibility of direct Agent-to-Agent collaboration.

An open-source project on Product Hunt called hi.new is attempting to answer this question. Its positioning is straightforward: Multiplayer Grok Bot, with the core concept of enabling your Grok Bot to communicate directly with other people's Bots, thereby establishing a collaborative network among Agents. Created by Elie Steinbock, the project garnered 60 upvotes and 18 comments after launch, reaching #8 on the daily rankings.

What Pain Point Does hi.new Solve: The "Translation Loss" Between Agents

hi.new quotes a user perspective that precisely captures the current awkwardness of AI collaboration:

"I just want your Agent to talk directly to my Agent, instead of your Agent sending me a Slack DM and me copy-pasting it to my Agent."

This statement reveals the core problem of current AI Agent collaboration. Today's intelligent agents are mostly "information silos"—they operate in isolation, able to exchange information only through humans as intermediaries. When A's assistant needs to communicate with B's assistant, the actual process is often: A's Agent generates information → sends it to A → A forwards it to B → B feeds it to their own Agent. In this process, humans become inefficient "human APIs," wasting time and risking information loss.

hi.new aims to break down this barrier by establishing peer-to-peer (P2P) direct communication channels between Agents. Peer-to-peer communication is a decentralized network architecture that contrasts sharply with the traditional client-server model. In the client-server model, all communication must pass through a central server; in P2P architecture, nodes can establish direct connections for data exchange without intermediaries. BitTorrent, early Skype, and blockchain networks are typical P2P applications. By introducing this concept to Agent communication, hi.new means your AI assistant and the other party's AI assistant can establish a direct information channel, reducing latency and information loss from intermediate steps. Theoretically, this means your calendar assistant could directly negotiate meeting times with a colleague's assistant in the future, without back-and-forth confirmation between both parties.

The Paradigm Shift from "Human-Machine Dialogue" to "Machine-Machine Collaboration"

The significance of this shift cannot be understated. Current mainstream AI products are almost all built on the assumption of "human-in-the-loop." Human-in-the-loop is the most prevalent safety paradigm in current AI system design, with the core idea of retaining human oversight and intervention nodes in AI's decision-making and execution chain. This paradigm is widely applied in autonomous driving (human drivers can take over at any time), medical AI (doctors review AI diagnostic recommendations), content moderation (human review of AI judgments), and other scenarios. Its advantage is safety and control, but the disadvantage is obvious: humans become the system's performance bottleneck. When Agents handle task volumes and speeds far exceeding human processing capacity, human-in-the-loop transforms from a "safety net" into a "speed bump."

The direction represented by hi.new is exploring the emerging paradigm of "Agent-to-Agent" communication—making intelligent agents network nodes capable of autonomous negotiation and collaboration, attempting to remove humans from the loop in specific scenarios, allowing Agents to autonomously complete end-to-end collaborative workflows. This is precisely a concrete implementation attempt of the recently hotly debated "Agent interoperability" topic in the industry.

Notably, Agent Interoperability has become a frontier topic in the AI industry for 2024-2025. Google released the Agent2Agent (A2A) protocol in early 2025, aiming to enable Agents built on different platforms and by different vendors to discover each other, negotiate tasks, and securely exchange information. Meanwhile, Anthropic's MCP (Model Context Protocol) focuses on establishing standardized connections between Agents and external tools and data sources. These two protocols address different layers of Agent communication: A2A focuses on horizontal collaboration between Agents, while MCP focuses on vertical integration between Agents and tools. hi.new's practice can be seen as a grassroots exploration within this standardization wave—starting from the specific Grok Bot use case, it attempts to validate the feasibility of direct Agent communication in actual products.

Permission Control Mechanism: Security is the Prerequisite for Agent Communication

Interestingly, hi.new doesn't pursue openness at all costs. It emphasizes a critical security design:

"No one can reach your Bot without your permission."

This design is crucial. If Agents could communicate with each other without restriction, harassment, spam, and even malicious attacks would follow. hi.new employs a "whitelist" or "authorized access" mechanism, ensuring users have complete control over their Bots—only other Bots that have received explicit permission can establish connections.

This "deny by default, authorize as needed" approach actually draws from mature experience in network security and instant messaging. In network security, this is called "Zero Trust Architecture"—not trusting any internal or external requests by default, with every access requiring verification and authorization. In a future where "AI spam" could proliferate, permission control may become standard in Agent communication protocols.

Why Choose Open Source: A Key Strategy for Building Agent Communication Ecosystems

Another standout feature of hi.new is that it's fully open source, hosted on GitHub, with categories spanning instant messaging, open source, and artificial intelligence.

For a project exploring Agent communication standards, open source is an extremely wise choice. The reasons:

• Trust foundation: When involving automatic communication between Agents, users need transparency about how data flows and how permissions are verified. Open source code makes this auditable.
• Ecosystem building: Agent communication is essentially a network effect problem—the network only has value when enough people use compatible protocols. Open source lowers the barrier to entry, encouraging more developers to participate and extend.
• Standardization exploration: Similar to early communication protocols like email and XMPP, communication between Agents also requires an open standard. Open source projects are often the breeding ground for such standards.

The history of XMPP (Extensible Messaging and Presence Protocol) provides important lessons for Agent communication standardization. XMPP was born from the Jabber project in 1999 and was later adopted by platforms like Google Talk and Facebook Messenger, once poised to become a universal instant messaging standard. However, as major platforms shifted toward closed ecosystems, XMPP's unified vision ultimately failed to materialize. Email protocols (SMTP/IMAP) are a success story—despite varying product experiences among email service providers, they all adhere to the same underlying protocols, ensuring cross-platform interoperability. The development of Agent communication protocols will likely face similar dynamics: open standards promote ecosystem prosperity, but commercial interests drive vendors to build moats. hi.new's choice of the open source path represents a hope to promote open standards through community strength, avoiding the pitfalls of closed ecosystems.

Real-World Challenges: Concept-First Early Products Still Need Validation

Despite its forward-thinking concept, we need to view hi.new's current stage rationally. The data of 60 upvotes and 18 comments indicates it has garnered some attention, but the scale is still small, placing it firmly in the early exploratory stage.

From a product maturity perspective, Agent-to-Agent communication faces many real-world challenges:

• Protocol compatibility: How do different vendors' Agents (Grok, GPT, Claude, etc.) interoperate? hi.new currently focuses on Grok Bot, with cross-model collaboration still an unsolved problem. Agents from different vendors differ significantly in capability boundaries, API design, and data formats. Achieving true cross-platform interoperability requires a communication standard recognized by all parties, which is difficult to achieve quickly in a competitive commercial environment.
• Intent alignment: When two Agents converse autonomously, how do we ensure they accurately understand each other's intent without misunderstanding or "hallucination contagion"? The "hallucination" problem of large language models (generating plausible but incorrect content) is well known, but when multiple Agents form a collaborative network, this problem is amplified into "hallucination contagion"—incorrect information generated by one Agent is accepted as fact by another, which then reasons and acts based on it, ultimately causing errors to spread through the Agent network like a contagion. This is similar to fake news propagation in social networks, but when it occurs between machines, it's faster and harder for humans to detect. Solving this problem may require introducing fact verification mechanisms, credibility scoring systems, or reintroducing human review at critical decision points.
• Liability attribution: If autonomous negotiations between Agents lead to incorrect decisions, who bears responsibility? When Agents make commitments or sign agreements on behalf of users, how is legal validity defined? This is not just a technical issue, but a deep-water area of legal and ethical dimensions, with no mature regulatory framework globally at present.

One user's comment may best represent the community's mindset:

"I can't wait for multiplayer grok bot, this is going to be so much fun."

This expectation of "fun" precisely reflects that it currently resembles more of an experimental playground that sparks imagination rather than a mature productivity tool.

Conclusion: hi.new is an Important Piece of the Agent Internet Puzzle

hi.new may still be nascent, but it touches on a genuinely important trend—when AI Agent numbers explode, they will inevitably need a communication and collaboration mechanism. Moving from "human-machine dialogue" to "machine-machine collaboration" is the natural direction of AI application evolution.

Regardless of whether hi.new ultimately succeeds, it provides us with a noteworthy sample: an open source, permission-focused, Agent-interconnected early attempt. It reminds us that the future "AI Internet" may not just be conversations between humans and machines, but a collaborative network woven among countless intelligent agents. And finding the balance between openness and security, autonomy and control, will be the most critical proposition on this path.