Week 6 slides: always on

← → to move, F for full screen, N for speaker notes

AI in Research · Week 6 · November 4, 2026

Always on

An agent that lives for days: what wakes it, how it remembers, how you stay in control, and how it survives a crash.

Parts 1–4 · from week 5

What wakes an agent, compaction and memory, budgets and the kill switch, crash and resume, running it as a service; then weeks 1–5 as one agent

Part 5 · moved to week 8

AI eval: we didn't reach it in class, so it opens the week 8 deck, at class.xhelio.ai/week-8

class.xhelio.ai/week-6

Week 6 · Recap of week 402

Last time: many agents

DesignWho decides the stepsTypical failure
One agentthe modelthe context fills up
Workflowyour codecan't handle the unexpected
Lead + subagentsthe lead modellossy reports, duplicated work
Peer teameach agent, plus a shared boardconflicting decisions, waiting forever

Pick the highest row that does the job. Today's agent is the first row, running for a week: "the context fills up" is Part 2.

Week 6 · Today03

Today: one agent's day

1 Waking up · what wakes it, push or poll, heartbeats, schedules, one wake-up end to end

2 Remembering · the context only grows, compaction on the wire, the cold cache, memory files

3 Control · approvals at 3 a.m., budgets, the kill switch, loops that never end

4 Surviving · crashes, resuming from the log, doing things once, running as a service

5 AI eval · moved to week 8: how models are trained and measured, and by whom

We didn't get to week 5 in class, so today starts there, at a faster pace.

Week 6 · Always on04

Always on is mostly asleep

asleep: waiting for something to happen 4 min6 min3 min9 min1 min2 min 07:0013:0019:0001:0007:00 An example day: running 25 min, asleep 23 h 35 min

Asleep costs nothing

No model calls, no tokens. A small Python process waits.

Something wakes it

A message, a timer, a heartbeat, another agent.

Then it's week 2's loop

It runs until it answers, then goes back to sleep.

Week 6 · Always on05

Always-on agents in 2026

AgentSinceRuns onWhat wakes it
OpenClaw (open source)Nov 2025your machine: one Gateway processchat messages, a heartbeat, scheduled jobs
Claude Code /loop, channelsMar 2026your open terminal sessiona recurring prompt; MCP servers that push messages in
Claude Code scheduled tasksApr 2026Anthropic's clouda schedule (hourly at most), an API call, an event
GitHub Copilot coding agentSep 2025GitHub Actionsan issue assigned to it
OpenAI Agents APISep 2026OpenAI's serversyour code; OpenAI runs compaction and recovery
Your harnessthis weekyour computer, as a serviceTelegram, email, a schedule

Why now: METR estimates frontier agents finish 12-hour tasks half the time (Claude Opus 4.6), and that length has doubled about every 4 months since 2023.

docs.openclaw.ai · code.claude.com/docs · github.blog, Sep 25, 2025 · OpenAI API changelog · metr.org/time-horizons (accessed Oct 23, 2026)

Week 6 · Always on06

OpenClaw, piece by piece

Telegram WhatsApp Slack, Signal, … the Gateway one long-lived process per host · sessions, one per chat· scheduled jobs (automations)· a heartbeat, every 30 min· transcripts on disk agent turnsmodel + tools, approvals for commands workspace files MEMORY.md · memory/2026-10-28.md HEARTBEAT.md
PartHow OpenClaw does it
Wakingheartbeat reads HEARTBEAT.md; a reply of HEARTBEAT_OK is dropped; jobs: at, every, cron, on-exit
Rememberingauto-compaction near the limit, after a silent turn to save notes; daily notes; nightly "dreaming" at 03:00
Control/stop also clears the queue; 3 identical tool errors end a turn; command approvals
Survivinga LaunchAgent or systemd unit that restarts it; sessions that were working resume after a crash

docs.openclaw.ai (architecture, memory, compaction, gateway, restart-recovery; accessed Oct 23, 2026) · heartbeat doc as of Jul 5, 2026 (github.com/openclaw/openclaw)

Week 6 · 1 Waking up07

An agent's day in five states

asleeprunningwaitingcrashedpaused an event arrives it answers · /stop asks to run a command you reply, or 10 minpass (= no) the process dies the service restarts it;it reloads the log /pause · /resume on your phoneonly the logsurvives
Week 6 · 1 Waking up08

What can wake it

Wake-upComes fromBecomes this user message
A messageTelegram, email, Slack[telegram] Jason: is the fit done?
A timera schedule you or it set[schedule daily-papers] It is 08:00. …
A heartbeatevery N minutes[heartbeat] Read HEARTBEAT.md. Anything due?
Another agenta subagent finishing[subagent notes] sa-2 is idle: …
The worlda job ends, a file appears[watch] run_042 finished, exit 0

Every wake-up ends up as text in the context. The model can't tell a timer from you unless the text says which it is.

Week 6 · 1 Waking up09

Watch in code, think in the model

Expensive

The model checks

"Every 5 minutes, check my email." That is 288 model calls a day, nearly all of them finding nothing.

Free

Your code checks

A thread asks IMAP for UNSEEN every minute. Telegram's getUpdates holds the line open until a message comes. Only something new wakes the model.

def watch_email(events): while True: for msg in imap.search("UNSEEN"): # a cheap question, no model involved events.put(Event("email", msg)) # this is what wakes the agent time.sleep(60)
Week 6 · 1 Waking up10

A heartbeat costs a call, even when nothing happens

A heartbeat every hour, gpt-6-astra, a 20,000-token prompt

$5 a day

24 calls × 20,000 tokens × $10 per million: after an hour asleep, assume the cache is gone. About $150 a month, mostly to hear "nothing new".

Check in code first. OpenClaw skips the call when HEARTBEAT.md is empty.

A small prompt. Its isolated heartbeat sends 2–5K tokens instead of ~100K of history.

A small model for heartbeats, the big one for work.

Say nothing. A reply of HEARTBEAT_OK is dropped, not sent to you.

OpenClaw heartbeat docs, as of Jul 5, 2026 (default every 30 min; "If nothing needs attention, reply HEARTBEAT_OK")

Week 6 · 1 Waking up11

A schedule: the clock lives in code

schedule.toml · read by the harness

[[job]] name = "daily-papers" cron = "0 8 * * 1-5" # 08:00, Monday to Friday task = """Search arXiv for yesterday's papers on switchbacks. Message me the three best, one line each.""" [[job]] name = "check-run" every = "2h" task = "Check ~/runs/042/log.txt; tell me only if it stopped."
0 8 * * 1-5

minute · hour · day · month · weekday

The model can add a job with a schedule tool ("every Monday, remind me…"), but the harness keeps the clock.

A model has no clock. It knows the time only if you tell it.

Week 6 · 1 Waking up12

One wake-up, end to end: a loop around the loop

events = queue.Queue() # Telegram, email, schedule, heartbeat start_watchers(events) # threads that put events here messages = load_log() # 4: carry on where it left off while True: event = events.get() # 1: asleep. No calls, no tokens if paused() or over_budget(): # 3: code decides, not the model hold(event); continue messages.append(user(event.as_text())) save(messages) # 4: on disk before acting run_loop(messages) # week 2's loop, until it answers if count_tokens(messages) > COMPACT_AT: compact(messages) # 2: make room

Week 2's loop is the inner loop. Always on adds an outer loop: wait, wake, work, save.

One thing never waits in the queue: /stop. The Telegram watcher acts on it at once (Part 3).

Week 6 · 2 Remembering13

A week-long conversation doesn't fit

context window without compaction: full by Thursday MonTueWedThuFri tokens with compaction: a sawtooth

Every wake-up adds messages, tool results and emails. Nothing ever leaves on its own.

Long before it's full, accuracy drops ("context rot"): at 32K tokens, 11 of 13 models fell below half their short-context score (NoLiMa, 2025).

So the harness must make room, and decide what to forget.

Modarressi et al., NoLiMa, ICML 2025 (arXiv 2502.05167) · Du et al., Findings of EMNLP 2025 (arXiv 2510.05381) · Chroma, Context Rot, Jul 2025

Week 6 · 2 Remembering14

Three ways to make room

WayWhat goesWhat you risk
Clear old tool resultsthe 400-line output from three days ago; the call stays, its result becomes "[cleared]"little: the agent can run it again
Summarize (compaction)the whole history, replaced by a summary the model writesthe summary drops a detail that matters later
Move it to files (memory)facts and decisions the agent writes to disk, read back when neededit forgets to write, or never reads them

Real harnesses use all three: clear first (cheap and safe), summarize near the limit, and keep files for what must survive.

Week 6 · 2 Remembering15

Compaction on the wire

Before: 170,000 tokens 240 messages: every tool result, every email, every detour compact at 150k systemtools One extra model call: "Summarize this conversation, so you can continue from the summary alone." Then the harness swaps the history for the summary, and keeps the last few messages. After: 21,000 tokens summary, 3,000 tokens last 6 messages, 8,000 tokens The same system prompt and tools stay at the front, so their part of the cache survives.
Week 6 · 2 Remembering16

What the summary must keep

You are about to lose this conversation. Write the summary you will continue from. Keep, under these headings: ## Goal what the person asked for, in their words ## Done what is finished, with results and numbers ## Open what is in progress, and the next step ## Decisions what was decided and why; what was ruled out ## Exact paths, run ids, DOIs, commands, error lines ## People who wrote, what they want, what you promised Copy identifiers exactly. Leave out tool output you can get again.

A summary is a prompt for your future self. Vague summaries lose the exact things: a path, a number, a promise.

Codex's begins: "You are performing a CONTEXT CHECKPOINT COMPACTION. Create a handoff summary for another LLM that will resume the task." In Claude Code you can add a focus: /compact keep the fit parameters.

Week 6 · 2 Remembering17

Asleep, the cache goes cold

Re-reading on waking (gpt-6-astra)Cache warm
$1 per million
Cache cold, hours later
$10 per million
150,000 tokens, never compacted$0.15$1.50 per wake-up
21,000 tokens, after compaction$0.02$0.21 per wake-up

Twenty wake-ups a day on the big context: $30 a day just to re-read it. A chat stays warm between messages; an always-on agent mostly wakes up cold.

The summary call itself is cheap: one read of the history plus a few thousand tokens out, about $0.30 on a warm cache.

Week 6 · 2 Remembering18

Who compacts: the harness, or the API

WhoWhenHow
Claude Codenear the window's end (200K, or about 967K on 1M models); or a big conversation left idle about 54 min, just before its cache expiresclears old tool output first, then summarizes; CLAUDE.md is re-read from disk; /compact <focus>; a PreCompact hook can block it
Codexat 90% of the window, unless you set model_auto_compact_token_limit"Create a handoff summary for another LLM that will resume the task"; replace it with compact_prompt
OpenAI APIat the compact_threshold you set (server-side, Feb 2026)an encrypted compaction item you can't read
Anthropic API150K input tokens by default (beta, Feb 2026); tool results cleared from 100K (Sep 2025)a compaction block; everything before it is dropped
Your harnessyour thresholdyour prompt, and a log of what it kept

code.claude.com/docs (how-claude-code-works, model-config, settings, hooks) · github.com/openai/codex (protocol/src/openai_models.rs, prompts/templates/compact) · developers.openai.com/api/docs/guides/compaction · platform.claude.com/docs (compaction-threshold, context-editing); accessed Oct 23, 2026

Week 6 · 2 Remembering19

Rules go in files, not in the chat

February 2026, a real inbox

A researcher tells her OpenClaw agent: suggest what to delete, don't act. The inbox is large; she believes it triggered compaction and the instruction was lost.

It starts deleting, and ignores the stop messages from her phone. "I had to RUN to my Mac mini like I was defusing a bomb."

In AGENTS.md or CLAUDE.md

Part of the system prompt, re-read from disk after every compaction. Never summarized.

"If an instruction disappeared after compaction, it was given only in conversation." (Claude Code docs)

The conversation is working memory. And her stop went through the model: a kill switch must not (Part 3).

TechCrunch, Feb 23, 2026 (TechCrunch could not independently verify what happened to the inbox) · code.claude.com/docs/en/memory

Week 6 · 2 Remembering20

Memory files: write it down before you forget

an agent's workspace

workspace/ ├── AGENTS.md # rules: in the system prompt, every call ├── MEMORY.md # long-term facts and decisions └── memory/ ├── 2026-10-27.md # yesterday's notes ← loaded at start └── 2026-10-28.md # today's notes ← loaded at start
[silent turn, near the limit] Session nearing compaction. Store durable memories now.

Before compacting, one extra turn to write notes (OpenClaw's "memory flush").

For long jobs, a progress file and git commits: each session is a new shift that reads the log first (Anthropic, Nov 2025).

A file can be searched and corrected; a summary can only be read.

docs.openclaw.ai/concepts/memory (and the Feb 14, 2026 version) · anthropic.com/engineering/effective-harnesses-for-long-running-agents (Nov 26, 2025)

Week 6 · 3 Control21

Who's in charge at 3 a.m.?

An email arrives. The agent wants to run python reduce.py. You are asleep.

PolicyWhat happensGood for
Wait for youthe task pauses until you replyanything that changes things
No answer means noafter 10 minutes the model is told "declined: nobody answered", and finds another way or stopsthe default for the night
Allowed in advancea short list runs without asking: ls, cat, one script in one folderread-only and routine work

Never "allow everything because I'm asleep". That is the night prompt injection arrives by email.

Week 6 · 3 Control22

Budgets are walls in code

limits.toml

turns_per_wake = 30 # then it stops and reports minutes_per_wake = 20 dollars_per_day = 2.00 # from each call's token usage replies_per_sender = 5 # per day, per address jobs_max = 10 # and none more often than hourly
[telegram → you] Daily budget reached ($2.00). Paused until midnight, or send /resume. Last task: summarizing 3 emails.

Checked by the harness before every model call. The model can't argue with it.

On a limit: stop, save, tell you. Never fail silently.

You already count tokens per call (homework 2). OpenClaw's docs list a 48-hour run limit and usage reports, but no dollar cap: that one you build.

Week 6 · 3 Control23

The kill switch is code, not a prompt

your phone/stop Telegram watcheryour chat id? a command? the harness actskill the process group, stop subagents reply"stopped python fit.py after 12 min" the model is never asked
CommandWhat the harness does
/statusstate, today's tokens and dollars, the next job, running subagents
/stopstops the current work now and goes back to sleep
/pause · /resumestops waking on events (messages wait) · starts again
Week 6 · 3 Control24

Loops that never end

Mail loop

Two agents, replying

Yours answers an auto-reply; the auto-reply answers yours. Thousands of emails by morning.

Retry loop

Paid, every minute

OpenClaw, fixed Oct 2026: heartbeat turns retried every minute while a chat channel was down.

Wake loop

Its own alarm clock

It schedules a job every minute "to check progress", and each job schedules another.

if msg["Auto-Submitted"] not in (None, "no") or msg["List-Id"]: # RFC 3834: never answer a robot return "skip" if replies_today(msg.sender) >= LIMITS.replies_per_sender: # and never more than 5 a day return "skip"

Every guard is a counter in code: replies per sender, turns per wake, jobs and their minimum interval.

Week 6 · 3 Control25

Always on means always exposed

Open doors

21,639 OpenClaw gateways reachable from the internet on Jan 31, 2026; over 135,000 by Feb 9.

Strangers' plugins

341 of 2,857 skills audited on its skill hub were malicious, most installing a password stealer.

Its inbox

A public challenge: 6,000 emails tried to inject an OpenClaw mail agent. None got the secret file; the bill passed $500.

An always-on agent is a server that reads strangers' mail: localhost only, no unknown plugins, updates, and remove a leg of the lethal trifecta.

Censys, Jan 31, 2026 · The Register, Feb 9, 2026 · The Hacker News, Feb 2, 2026 (Koi Security audit) · Decrypt, Jun 26, 2026

Week 6 · 4 Surviving26

Processes die

  • The lid closes; the laptop sleeps.
  • Wi-Fi drops; the API answers 429 or 500.
  • Out of memory; a bug; one watcher thread dies quietly.
  • You close the terminal; the OS restarts for an update.

Only what's on disk survives. Plan for the restart, not against the crash.

The sneaky one

One watcher thread dies with an exception. The process is still running and the agent looks alive, but it never wakes again.

So keep the control path, your phone commands, separate and simple: it has to work when the rest is broken.

Week 6 · 4 Surviving27

Save every step, resume from the log

logs/main.jsonl · one line per message, appended, flushed

{"t":"03:12:09","role":"user","content":"[email] from lingtai2026@…"} {"t":"03:12:14","role":"assistant","tool_calls":[{"id":"c1","name":"bash", "arguments":{"command":"python reduce.py run_042"}}]} — the process dies here —
messages = [json.loads(line) for line in open(LOG)] if unanswered_tool_call(messages[-1]): # died mid-command messages.append(tool_result("Interrupted by a restart; not re-run. " "Check its output before trying again."))

Append, never rewrite. A crash can cut off the last line, not the history.

On restart, the last line says where it was. Don't blindly re-run what was in flight: tell the model, and let it check.

Compaction appends a marker line with the summary; on restart, read from the last marker.

Week 6 · 4 Surviving28

Do it once

03:12 sends the reply to lingtai2026 ✓ sent 03:12 — crash, before the log line is written — 03:13 restarts, sees an unanswered call, sends it again ✗ twice
def send_email(to, subject, body): key = digest(to, subject, body) # the same email, the same key if outbox.has(key, "done"): return "already sent" if outbox.has(key, "intent") and sent_folder_has(key): outbox.mark(key, "done"); return "already sent" outbox.mark(key, "intent") # 1. what you're about to do smtp_send(to, subject, body, message_id=key) # 2. do it outbox.mark(key, "done") # 3. that it's done

Reading is safe to repeat. Sending, deleting, paying are not.

Give each action a key, write the intent first, and on restart check before redoing.

Payment APIs call this an idempotency key.

Week 6 · 4 Surviving29

Run it as a service

macOS · ~/Library/LaunchAgents/edu.ucla.myagent.plist

<key>ProgramArguments</key> <array> <string>/Users/me/harness/.venv/bin/python</string> <string>/Users/me/harness/agent.py</string> <string>--always-on</string> </array> <key>RunAtLoad</key> <true/> <key>KeepAlive</key> <true/> <!-- restart if it dies --> <key>StandardErrorPath</key> <string>/Users/me/harness/logs/agent.err</string>

Linux · ~/.config/systemd/user/myagent.service

[Unit] Description=my research agent [Service] WorkingDirectory=/home/me/harness ExecStart=/home/me/harness/.venv/bin/python agent.py --always-on Restart=on-failure # restart if it dies RestartSec=10 [Install] WantedBy=default.target

The OS starts it, restarts it, and keeps its logs. tmux survives closing the terminal, not a reboot. A Mac with the lid closed sleeps, and so does your agent: keep it plugged in and awake (caffeinate -i).

Week 6 · 4 Surviving30

A heartbeat you can see

status.json · rewritten every minute

{ "state": "asleep", "since": "2026-10-28T03:14:09-07:00", "last_event": "email from lingtai2026@gmail.com", "today": {"wakeups": 14, "tokens": 412318, "dollars": 0.61}, "next_job": "daily-papers at 08:00", "subagents": 0, "pid": 48211 }

/status on your phone reads it.

A watchdog checks its age: older than 5 minutes means the agent is stuck or dead, so it tells you.

A daily digest at 9 p.m.: what it did, what it spent, what failed.

Week 6 · Wrap-up31

The whole picture

ProblemWhat it takesWho enforces it
Waking upwatchers in code, one event queue, a schedule file, labelled wake-up messagesthe harness
Rememberingclear old results, compact near the limit, rules in AGENTS.md, notes in memory filesthe harness, and the model's writing
Controlbudgets, approvals that time out to no, commands from your phone, loop guardsthe harness, never the model
Survivingan append-only log, resume without re-running, actions done once, a service, a watchdogthe harness and the OS

Almost none of it is the model. An always-on agent is a small, careful program that calls a model when something happens.

Weeks 1–5 · Wrap-up32

The anatomy of an agent

Only the orange box is a trained network. The rest is code you can read, change and test, and only code enforces.

Weeks 1–5 · Wrap-up33

Loops inside loops

Each loop runs inside the next: tokens in milliseconds, tool calls in seconds, wake-ups over days.

Weeks 1–5 · Wrap-up34

One email at 3 a.m., through every box

TimeWeekWhat happens
03:12:005A watcher sees a new email. Code checks pause and budget, and logs it. No model yet.
03:12:013The harness builds the context: rules, AGENTS.md, skill list, tool schemas, memory, and the email as untrusted text.
03:12:012One stateless call sends all of it. After hours asleep, the cache is cold: full price.
03:12:041The model predicts tokens. They spell a tool call: search the literature.
03:12:054A subagent searches in a fresh context; a ten-line report comes back.
03:13:402The report is appended, the loop calls again, and the model writes a draft reply.
03:13:413Sending mail needs your approval. Nobody answers within ten minutes, so the answer is no.
07:30:005You read the night's summary on your phone and approve by reply. The email is sent once, and logged.

About 100 seconds of model time. Everything else was your code, waiting.

Weeks 1–5 · Wrap-up35

Every box is paid for in tokens

WeekWhat runs up the billWhat brings it down
1The model you pick. The same capability gets about 13× cheaper every year.The cheapest model that passes your check, and ask again in a few months.
2Each call resends everything. Output costs about 5× input.Caching: a cached read is 1/10 to 1/50 of the price. Stable things first. One run: $19 → $5.
3Every tool schema rides along: five MCP servers, 55,000 tokens before you type.Names in the context, details on demand. Tool search: 85% fewer tokens.
4Many agents: several times the tokens of one.Only for breadth, or for an independent check.
5Waking up cold: $10 instead of $1 per million. An hourly heartbeat: about $5 a day.Watch in code, compact, and give heartbeats a small prompt and a small model.

The same rule five times: send the model less, and call it less often.

Week 6 · Homework36

Homework 5: keep your agent on

  1. Always on: a service that restarts it; wakes on Telegram, email and a schedule, watched in code.
  2. Remembering: compaction at a threshold, tokens before and after; AGENTS.md and a memory file loaded at every wake.
  3. Control: limits in code; /status /stop /pause /resume from your phone; no answer means no; never reply to a robot.
  4. Surviving: an append-only log; it survives kill -9, never re-runs blindly, never sends twice.

Show it

Leave it on overnight; turn in the log.

Five demos: compaction, /stop, a budget, a crash, the night.

failures.md: everything that went wrong. Next slide: the first ones in.

No harness of your own? Wrap Claude Code or Codex in a small always-on loop. Due Wed Nov 4 · class.xhelio.ai/week-5/homework

Week 6 · Your failures.md37

What your agents did overnight

PartWhat happenedCaught by
RememberingA memory write returned saved=true. The file now held a 197-character heading; the good notes were gone.opening the file
RememberingPowerShell wrote MEMORY.md as UTF-16; the reader assumed UTF-8. Compaction failed, then the next wake shut the agent down.the log
WakingTwo inbox tools returned only the newest 20 headers. A burst of mail would never be seen.reading the code
WakingThe scheduled test ran, and the model copied its instruction into the reply: "Do not use tools."reading the reply
ControlOne failed compaction: 65,412 tokens in three calls, against a limit of 100,000 a day. The limit held.the daily counter
SurvivingRestart on failure, set to one minute: exit code 1, and no restart within 100 seconds.a deliberate crash

Each was found by checking, and the first even passed its own check. So how do you check an AI at all? That is week 8.

failures.md in homework pull requests #30, #31 and #32, as of Nov 3, 2026

Week 6 · Homework38

Homework 6: make your task hold

  1. Read what happened: your task's page and the agent's full trace. What was the decisive move?
  2. Fix its solvability: a new session of your agent reviews it with the check-solvability skill.
  3. Harden it: withhold the truth, use real data, ask for hidden causes, chain the steps. More of the same doesn't work.
  4. Review it again, then open your pull request. We run GPT-5.6 Sol on it.

The goal

GPT-5.6 Sol doesn't solve it.

A human expert would need over a year, ideally ten.

It must stay solvable: a right answer exists, and your verifier knows it. Still solved at the deadline? That still counts.

Update your task in week-1/submissions/. Due Wed Nov 18 · class.xhelio.ai/week-6/homework

Encouraged: propose it to Terminal-Bench-Science by Dec 15. One merged task qualifies you for co-authorship.

Week 6 · Wrap-up39

Takeaways

Always on

  • It's mostly asleep: watch in code, wake the model only when something happens.
  • Compaction trades detail for room. Rules go in files; it wakes to a cold cache.
  • Control is code, not a prompt: budgets, timeouts to no, a kill switch, loop counters.
  • Only the log survives a crash: append, resume without re-running, do each action once.

Moved to week 8

AI eval: how models are trained, how they are measured, and by whom.

Start there with the question your failures.md raised: how do you check an AI at all?

class.xhelio.ai/week-8

Homework 6: class.xhelio.ai/week-6/homework · Reading: class.xhelio.ai/week-5/reading · Still to come: how AI fails; keep your failures.md.