Homework · Fall 2026
Every week's homework builds on one of two things you keep for the whole quarter. Both are submitted as pull requests to the homework repository.
bash. It gains more tools, MCP servers, skills and teammates as the course goes on.
Next · due before class, Wednesday, November 18
Go back to your week 1 task. Fix its solvability with a review from a fresh session of your agent, 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.
Open an AI coding agent in your clone of the homework repository and give it this:
Read https://class.xhelio.ai/week-6/homework.md and help me with my homework.
Week 2 · due October 14 · results are in
Write the loop from class yourself: your program sends the conversation to a model, runs the shell commands it asks for (after you approve each one), and sends back the results. One tool, bash; a chat in a local web page or the terminal. Your AI coding agent writes most of the code with you. You submit a link to your harness's public repository and a skill that installs it.
Open an AI coding agent in a folder under your home directory and give it this:
Read https://class.xhelio.ai/week-2/homework.md and help me with my homework.
Your task · week 1
Frontier AI agents attempt every task, alone, in a sealed container, graded by your own verifier. Each week you use what their attempts expose to make the task harder and fairer.
Week 1 · round 1 · Claude Opus 5.5, xhigh effort
Open a finished task for the full analysis: what the agent did, why it passed or failed, and how to make the task harder or fairer. Reward is the official all-or-nothing grade from your verifier; partial score is our own regrade, which gives credit for being close.
| Task | Result | Reward | Partial score | Agent time |
|---|
One more submission, tumbling-satellite-telemetry, is being revised: its reference solution doesn't run yet, so the agent hasn't attempted it.
Your task, over the quarter