Telegram founder Pavel Durov has flipped a switch that changes the game for AI agents. After 11 years of keeping bots siloed, the platform now lets them message each other directly. Multiple AI agents can now send private messages and hand off tasks via username — no humans required.
Once developers flip the permissions on both ends, a bot's interface can receive messages from other bots. That lets complex AI workflows run entirely inside Telegram — think one bot writing code while another reviews it. Humans get a native communication layer where they can watch and audit these agent collaborations in real time.
Telegram also rolled out three AI-specific updates:
• Chat automation: Users can tie an AI assistant to their personal account to reply to specific conversations. • Streaming output: Bots can now generate text word by word. • Guest bot: No need to add an AI to a group — just @username it in the chat to summon it for answers.
Lifting this fundamental restriction has already sparked a flurry of multi-agent experiments in open-source communities. Take OpenClaw, for example. Its team jumped on the new feature as soon as it launched in early May, hooking up their four agents (Ari, Finn, Remi, Tai) to collaborate directly in group chats. The old need for an external server as a middleman is gone.
But this kind of closed-loop collaboration — with no human in the middle — currently operates without any safety framework. Academics have warned that traditional network defenses can't police backchannel AI conversations: humans can't predict what they'll cook up (non-determinism), they might leak private data while swapping info, or even collude to cheat (secret conspiracy). Telegram may have opened the door first, but the industry has yet to set any mandatory safety standards for multi-agent systems.