On May 7, 2026, Telegram became the first billion-user messaging platform to enable native, direct communication between autonomous AI bots \u2014 a capability that developers and AI researchers have sought for years and one that lands well ahead of any enacted federal standard for securing the multi-agent systems it makes possible.<\/p>\n
The update<\/a>, announced by founder and CEO Pavel Durov in a post that also introduced more than ten other new features, expands the Telegram Bot API to allow what the company calls “bot-to-bot communication mode.” Under the new system, one bot can send a private message directly to another bot by referencing its @username \u2014 but only when both the sender and the recipient have explicitly opted into the mode. That mutual opt-in is a deliberate design choice intended to prevent automated spam chains.<\/p>\n The change removes a technical barrier that had existed since the Bot API launched more than a decade ago. Until May 7, Telegram bots could not address each other directly; developers who needed multi-agent workflows had to route messages through a custom server acting as an intermediary. That workaround is now unnecessary.<\/p>\n Telegram’s Bot API Gains a Native Multi-Agent Coordination Layer<\/p>\n Telegram outlined several concrete use cases in its announcement: a code-review bot receiving a request from a collaborator bot and returning results to a human-monitored group chat; enterprise booking and customer-service bots delegating sub-tasks to specialist bots; and multi-step AI workflows that execute end-to-end without any human relay point. In each scenario, users who choose to watch can observe bot-to-bot conversations, making Telegram’s infrastructure function as an observable coordination layer rather than a black-box backend.<\/p>\n The practical significance is that Telegram’s Bot API already hosts more than 10 million bots<\/a>. Those bots can now form networks. A developer deploying an AI research agent, for example, can have that agent offload a data-retrieval subtask to a dedicated specialist bot and receive the result back \u2014 all within Telegram’s native message delivery infrastructure, using an audience of more than one billion existing users as a deployment environment.<\/p>\n The same update introduced Chat Automation, which lets individual Telegram users connect a bot to their own profile to handle replies on their behalf, with user-configurable controls over which conversations the bot can access. Bots also now support Streaming Text, displaying responses progressively as a model generates them. A Custom AI Styles feature lets users write prompts that define writing voices, shareable via links.<\/p>\n Late March’s Managed Bots Feature Set the Stage; Developers Responded the Same Day<\/p>\n The bot-to-bot announcement builds directly on Managed Bots, introduced in the March 31, 2026<\/a> update. That feature lets a single “manager bot” create and configure personalized agent bots for individual users in two taps, eliminating the manual BotFather token setup. The combination \u2014 Managed Bots for frictionless deployment, bot-to-bot communication for agent coordination \u2014 constitutes a coherent agentic platform within a consumer messaging app.<\/p>\n Developer response was immediate. On May 7, the team behind OpenClaw, an open-source AI agent runtime that uses Telegram as its primary communication layer, filed a GitHub issue<\/a> the same day requesting implementation support for the new feature. The issue describes a four-agent setup \u2014 agents named Ari, Finn, Remi, and Tai \u2014 in which one agent would delegate a billing query to a specialist peer and surface the answer to a human, with no human chat thread serving as an intermediary relay. Before May 7, the team had built a workaround routing system because Telegram bots could not see each other; the new API removes that constraint at the platform level.<\/p>\n Multi-Agent Security Research Identifies Risks the Platform Does Not Yet Address<\/p>\n The same week Telegram shipped the feature, the security community was actively publishing on the risks of exactly these kinds of systems. A March 2026 study<\/a> by researchers at Georgia Institute of Technology, Kennesaw State University, and the OWASP Foundation surveyed seventeen existing security frameworks against multi-agent system cybersecurity risks and found that the best-covered framework \u2014 the OWASP Agentic Security Initiative \u2014 addressed only 65.3 percent of the identified threat categories. Non-determinism and data leakage were identified as the two most under-addressed risk domains across all frameworks reviewed.<\/p>\n