Week 5
Always on
Wednesday, October 28, 11:00–11:50 a.m., Slichter 3853
Joining remotely? The Zoom link is on the home page.
Your agent now has a phone line, an inbox and helpers. This week we ask what it takes to leave it running for a week. An always-on agent is mostly asleep: what wakes it, and how does it avoid paying for doing nothing? How does it keep working once its conversation no longer fits, and what does it forget? How do you stay in control while you sleep? And what happens when the process dies at 3 a.m. halfway through sending an email?
In class
Four parts, one for each problem an agent meets only when it stays on. Each card opens the slides at the start of that part; the reading page has the details. The class ends with a wrap-up of weeks 1–5: one agent, box by box, what each box costs, and where it breaks.
- Part 1Waking upAn agent's five states, what can wake it, watching in code instead of the model, heartbeats, schedules, and the loop around the loop.
- Part 2RememberingThe context only grows: three ways to make room, compaction on the wire, the cold cache, and memory files.
- Part 3ControlApprovals at 3 a.m., budgets in code, the kill switch, loops that never end, and what an always-on agent exposes.
- Part 4SurvivingProcesses die: resuming from the log, doing each action once, running as a service, and a heartbeat you can see.
Further reading
The details behind every slide, with sources: how OpenClaw does it, compaction in Claude Code, Codex and the APIs, the evidence on long contexts, the service files in full, and spacecraft that did all of this first.
Homework 5
Keep your agent on: it runs as a service, wakes on Telegram, email and a schedule, compacts its conversation, stays inside limits you can enforce from your phone, and survives a crash. Leave it on overnight and write down what went wrong. Due before class on November 4.