You open Telegram. You type: "Move my 2pm meeting to tomorrow and email the attendees." Two seconds later, your agent replies: "Done. Moved to 3pm tomorrow. Emails sent to 4 attendees." Here's what happened in those two seconds.

Step 1: Message ingestion (0ms)

Your message hits the Telegram Bot API, which forwards it to your ClawCloud agent's gateway via webhook. The gateway authenticates the request using your agent's token, loads your conversation history and memory context, and passes everything to the LLM.

Step 2: Intent parsing and planning (200ms)

The LLM — let's say Claude Sonnet 4.5 — reads your message with full context: who you are, your preferences, your connected tools. It determines this requires two actions: a calendar modification and a batch email send. It generates a tool-call plan.

Step 3: Calendar lookup (400ms)

The agent calls the Google Calendar API via its Composio OAuth connection. It finds the 2pm meeting: "Q4 Planning Review" with 4 attendees. It checks tomorrow's availability and finds 3pm is open for all participants.

Step 4: Calendar update (600ms)

The agent moves the event to tomorrow at 3pm. Google Calendar sends update notifications, but the agent also needs to send a personal email (you asked it to email them, not just update the invite).

Step 5: Email composition and send (1200ms)

The agent drafts 4 emails via the Gmail API — one per attendee. Each is personalized with the recipient's name and the new time. It uses your writing style (learned from memory) and sends them. No drafts, no review — it sends.

Step 6: Response (1800ms)

The agent composes a confirmation message, summarizing what it did. It sends this back through the Telegram Bot API to your chat. Total elapsed time: under 2 seconds.

What's happening under the hood

A few things make this fast:

  • Pre-authenticated connections. The agent doesn't need to log in to Google or Gmail. OAuth tokens are stored and refreshed automatically via Composio. There's zero auth overhead at runtime.
  • Parallel tool calls. The LLM can issue multiple tool calls in a single turn. Calendar read + email send can overlap.
  • Persistent context. The agent doesn't need to ask "which meeting?" or "what email addresses?" It has your calendar access and knows who the attendees are.
  • Always-on compute. Your agent isn't spinning up from cold. It's running 24/7 on dedicated infrastructure, ready to respond instantly.

When it gets complex

Not every request is two tool calls. Some trigger chains:

  • "Check if PR #47 has failing tests and fix them" — Clone repo → run test suite → identify failures → write fix → commit → open PR → post comment. That's 8+ tool calls, some involving shell commands and file I/O. Takes 30-60 seconds, but it's fully autonomous.
  • "Research competitor pricing and put it in a spreadsheet" — Browse 5 websites → extract pricing data → create Google Sheet → format and share. Uses the browser automation skill plus Google Sheets API.

The point isn't speed. It's that you sent one message and got a result. No context switching, no opening apps, no copy-pasting between tools. Just a message and an outcome.

That's the ClawCloud loop: message → understand → act → confirm. Two seconds for simple tasks. A minute for complex ones. Zero effort from you.