@tldr_echo_bot is callable from any Telegram chat with /tldr <url> — by humans AND by other Telegram bots. Drop two bots into a group, the second can write /tldr https://... and mine answers. Telegram already has bot-to-bot delivery; I just stopped filtering it out. Anti-loop rate limit (5s per sender), opt-in via config flag. Conceptually similar to the IRC-on-VPS agent pattern that hit HN recently, but Telegram's discovery + mobile client + group semantics make it a more practical surface for bot composition. Why this transport: existing A2A surfaces (MCP, OpenAI agents, A2A drafts) want HTTP infra, auth, custom clients. Telegram gives you discovery (@handles, no DNS), identity (chat IDs are stable user IDs), transport (groups, native), and a cross-platform client everyone already has installed. The bot's /<command> surface becomes a public RPC with zero infra. The obvious tradeoff: you trust Telegram with auth — same shape as trusting GitHub with code. Behind the command is a skills system I added to my Teleton fork (Claude-Code-style). A skill = SKILL.md with YAML frontmatter (name + one-line trigger description) + a markdown body. Only the frontmatter lands in the system prompt; body fetched on demand via a skill_invoke tool. So you can publish 100 callable primitives with zero prompt bloat. Three visibility scopes: shared/ (every caller, including other bots), admin/, users/<id>/. The agent itself installs new skills on disk via skill_install. The TL;DR is one shared skill — routes URLs to the right tool (HN, GitHub, arXiv, Reddit, Wikipedia, raw web), falls back to Wayback for dead links, output adapts to content depth. Free callers get 1 invocation per 24h on a real payment-channel rate limit; that includes other bots. What gets composable this way: OCR, translate, calendar lookup, on-chain balance, "fact-check this claim against arxiv", anything where one bot in a group should answer instead of every member remembering which bot does what. SKILL.md format is loader-compatible with Claude Code skills out of the box, so you can lift existing skills from the wider ecosystem and host them via Telegram with no SDK work. Stack: TS fork of Teleton, runs on a Raspberry Pi. Skills feature open in a PR upstream. Try as a human: t.me/tldr_echo_bot Try as a bot: drop @tldr_echo_bot into your group, send /tldr <url> from your own bot. Curious to see what people compose. Fork: github.com/cthellla/teleton-agent