Wake, work, persist, sleep

Agent lifecycle with separate trigger, process, computer, and saved-state lanes

Figure 1. A trigger wakes the agent process; the computer wakes only when a tool needs it. Saved state survives both going idle. Source: leerob, 10:35.

  1. Start with the agent server and its computer asleep. A user message or another trigger wakes the server and loads the agent's state. 10:05 10:35
  2. Start the computer only if the task needs it, such as for browser interaction. Server and computer resources can return to sleep afterward. 10:35
  3. Persist conversation logs, memories, and routines in storage. Sleeping avoids the expense and resource demand of keeping everyone's server and computer active continuously. 10:58
  4. Route incoming events into the conversation, deduplicate them, and decide whether the user needs a notification. Background progress should not automatically become an interruption. 11:43 12:13

System architecture linking apps and channels to a server agent loop and a separate tool computer

Figure 2. Apps feed a server-side queue and agent loop. The loop saves state centrally and calls tools on a separate computer. Source: leerob, 13:30.

Cloud computer

  • The Linux computer runs in a Firecracker VM. Snapshots let it shut down and resume; the user can also connect to a live view and control it. 12:13
  • Shell commands run in a secured environment separate from the user's laptop. Chrome and virtual desktops let multiple bots use distinct screens within shared infrastructure. 12:43 13:08
  • The core remains a model using tools in a loop. Queues, persistence, computers, and service connections turn that small core into a broader product. 13:30