Week 6

Always on

Wednesday, November 4, 11:00–11:50 a.m., Slichter 3853

Joining remotely? The Zoom link is on the home page.

We didn't get to week 5 in class, so this week starts there: what it takes to leave your agent running for a week. What wakes it, how it keeps working once its conversation no longer fits, how you stay in control while you sleep, and what happens when the process dies at 3 a.m. We didn't reach the planned second half, AI eval, in class: it moved to week 8.

In class

Four parts, at a faster pace than usual. Each card opens the slides at the start of that part. Part 5, AI eval, moved to week 8.

  1. Part 1Waking upAlways on is mostly asleep: what can wake an agent, watching in code, heartbeats, schedules, and the loop around the loop.
  2. Part 2RememberingThe context only grows: three ways to make room, compaction on the wire, the cold cache, and memory files.
  3. Part 3ControlApprovals at 3 a.m., budgets in code, the kill switch, loops that never end, and what an always-on agent exposes.
  4. Part 4SurvivingProcesses die: resuming from the log, doing each action once, running as a service, and weeks 1–5 as one agent.
  5. Part 5 · in week 8AI evalNot reached in class. Pre-training, post-training, reinforcement learning, Harbor, who does the measuring and how to read a score now open the week 8 slides.

Further reading

For the always-on agent, the week 5 reading page has it all, with code you can copy. The AI eval reading moved to week 8 with its slides.

Homework 6

Make your task hold: fix your week 1 task's solvability, then harden it until GPT-5.6 Sol fails on the science. Aim for a task that would take a human expert more than a year, ideally ten. Due before class on November 18. You're encouraged to take it on to Terminal-Bench-Science, where a merged task earns co-authorship.