How OpenClaw Works: The Architecture Behind the 'Magic'

OpenClaw is fundamentally an agent runtime with a gateway in front of it (2:00). The gateway is a long-running process that accepts connections from messaging apps (WhatsApp, Telegram, Discord, iMessage, Slack) and routes messages to AI agents. It doesn't think or reason—it simply accepts inputs and routes them to the right place.

The system accepts five types of input that, when combined, create the illusion of autonomy (2:45). These are: messages from humans, heartbeats from timers, cron jobs on schedules, hooks from internal state changes, and webhooks from external systems. There's also a bonus input type: agents can message other agents.

Heartbeats are simple timers that fire every 30 minutes by default, sending a preconfigured prompt to the agent (3:50). This is described as the "secret sauce" that makes OpenClaw feel proactive—the agent keeps acting even when you're not talking to it, but it's just responding to timer events with pre-written instructions like checking email or reviewing calendars.

Crons provide more precise control than heartbeats, allowing users to specify exact times and instructions (5:00). Examples include checking email at 9 AM daily, reviewing calendars every Monday at 3 PM, or browsing Twitter at midnight. The viral example of an agent texting someone's wife was simply a cron job firing at 8 AM and 10 PM with "good morning" and "good night" prompts.