Week 5 · Further reading
Always on, in more depth
The slides covered the main ideas. This page has the details and the evidence, with sources, checked on October 23, 2026. Always-on products change weekly (OpenClaw almost daily): when this page and a product's docs disagree, trust the docs.
- Always-on agents in 2026
- Spacecraft did it first
- OpenClaw, in more depth
- Waking up: watchers, heartbeats and schedules
- Compaction in Claude Code, Codex and the APIs
- Context rot: the evidence
- Memory files and long-running harnesses
- Control while you sleep
- Always on, always exposed
- Surviving: the log, doing things once, and the service
- When nobody was watching
1. Always-on agents in 2026
| Agent | Since | Where it runs | What wakes it |
|---|---|---|---|
| OpenClaw (open source, MIT) | Nov 2025 | one Gateway process on your machine or a server | chat messages (WhatsApp, Telegram, Slack, Discord, Signal, iMessage, …), a heartbeat, scheduled jobs, webhooks |
Claude Code /loop | Mar 2026 | your open session | a prompt repeated on an interval (1 minute minimum; a loop expires after seven days) |
| Claude Code channels (research preview) | Mar 2026 | your open session | MCP servers that push messages in: Telegram, Discord, iMessage. "Events only arrive while the session is open" |
| Claude Code scheduled tasks (formerly routines) | Apr 2026 | Anthropic's cloud: "nothing depends on your laptop being open" | a schedule (hourly at most), an API call or an event |
| GitHub Copilot coding agent | Sep 2025 (GA) | GitHub Actions | an issue assigned to it; it opens a draft pull request |
| Cursor background agents (now "cloud agents") | May 2025 (preview) | Cursor's cloud | a task you hand it |
| OpenAI Agents API (public beta) | Sep 2026 | OpenAI's servers, a managed Codex harness | your code; OpenAI "handles session orchestration, context compaction, and recovery" |
Why now: METR measures the length of task, in human working time, that an agent finishes half the time. Its latest data (Time Horizon 1.1) puts Claude Opus 4.6 at about 12 hours (95% interval 5–61 hours), and since 2023 that length has doubled about every 129 days. METR says measurements above 16 hours are unreliable with its current tasks, and stopped updating the page in September 2026. A twelve-hour task doesn't fit in one sitting; it fits in a night.
OpenClaw docs · Claude Code: scheduled tasks · Claude Code: channels · Introducing routines · GitHub, Sep 25, 2025 · Cursor 0.50 · OpenAI API changelog · METR time horizons (all accessed October 23, 2026)
2. Spacecraft did it first
Deep-space missions have run "always-on agents" for decades, with a delay no chat app has: New Horizons' signal took 4 hours 26 minutes to reach Earth when it woke for Pluto, Voyager 1's takes 22.5 hours. Their operating habits map almost one to one onto this week.
- Hibernation. New Horizons spent 1,873 days in hibernation on its way to Pluto, "about two-thirds of its flight time", in 18 periods from mid-2007 to late 2014 lasting 36 to 202 days. Most systems were off; the radios, main computer, power distribution and thermal control stayed on, and the spacecraft spun at 5 rpm instead of pointing. Hibernating reduced wear on the electronics, cut operations costs and freed Deep Space Network time. In 2025–26 it hibernated for a record 321 days and woke in June 2026: "Every status report through this hibernation period was 'green'".
- A beacon, not a conversation. Each week the computer sent one beacon tone: "a 'green' coded tone if all is well, or … one of seven coded 'red' tones if it detects a problem and requires help" (2006 press kit; a 2017 blog post counts seven tones in all). Engineering telemetry came about monthly. The idea was tested on Deep Space 1, whose beacon picked one of four tones, from "fine, and it does not need contact" to "red alert".
- Stored commands. New Horizons implements "almost all on-board activities" as time-tagged rules that fire when mission elapsed time passes a value (storage for 512), and autonomy rules that fire when a telemetry expression has been true often enough. Its final wake-up, on December 6, 2014, came from pre-programmed commands. Command loads are double-buffered, and a rule checks that the next load is in place; if it isn't, the spacecraft goes safe and points at Earth.
- Heartbeats and watchdogs. Fault protection watches "a pulse per second 'heartbeat'" from the active computer; after 180 seconds without it, the backup computer is powered on and takes over. A command-loss timer is reset to, for example, a week by every command from Earth; if it reaches zero, the spacecraft assumes its receiver failed and switches hardware.
- Safe mode. JPL's fault-protection practice: put the spacecraft "in a safe, commandable state which can be maintained for a reasonable period (typically two weeks)"; "the normal stored sequence is terminated and non-essential spacecraft loads are powered off"; priorities are hardware safety, then uplink, then downlink; and "fault responses should not autonomously resume the terminated sequence". The same rules make a good agent: stop the work, stay reachable, and let a human decide what to re-run.
- A fault in the schedule. On July 4, 2015, ten days before the Pluto flyby, New Horizons switched to its backup computer and entered safe mode. No hardware or software had failed: the cause was "a hard-to-detect timing flaw in the spacecraft command sequence", reported as a large command load arriving while the computer was compressing science data. Science resumed on July 7.
- Remote recovery. Voyager 1 began sending unreadable data on November 14, 2023. A memory readout showed about 3% of its flight data subsystem's memory corrupted, probably by one failed chip. With 45 hours per command and reply, engineers split the affected code into pieces and moved it elsewhere in memory, updating every reference; engineering data came back on April 20, 2024, and all four instruments by June 13. It worked because the spacecraft kept receiving commands throughout.
APL, Nov 13, 2014 · APL, Dec 8, 2014 · New Horizons launch press kit (2006) · Fountain et al., The New Horizons Spacecraft (2008) · NASA, Jul 7, 2026 · JPL: DS1 beacon monitor · Basics of Space Flight, ch. 11 · NASA PD-ED-1243, Fault Protection · NASA, Jul 6, 2015 · NASA Voyager blog, Apr 4, 2024 · JPL, Apr 22, 2024 · JPL, Jun 13, 2024
3. OpenClaw, in more depth
OpenClaw is "a self-hosted gateway that connects your favorite chat apps" to AI agents, and the always-on agent most people have heard of: about 393,000 GitHub stars in October 2026, more than React. It began in November 2025 as Clawdbot, was renamed Moltbot on January 27, 2026 after Anthropic asked for a name change, and became OpenClaw on January 30. Its creator, Peter Steinberger, joined OpenAI in February and said the project would move to a foundation. It changes almost daily, so treat what follows as a snapshot.
- One process. "A single long-lived Gateway owns all messaging surfaces"; one per host, listening on
127.0.0.1:18789by default. All session state belongs to the Gateway: direct messages share one session, each group chat gets its own, and each scheduled job runs in a fresh one. - Heartbeat (the classic design, January–July 2026). Every 30 minutes by default, the Gateway sends a fixed user message: "Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats. If nothing needs attention, reply HEARTBEAT_OK." A reply that is just
HEARTBEAT_OKis dropped. IfHEARTBEAT.mdis effectively empty, the call is skipped "to save API calls". The docs warn that "Heartbeats run full agent turns", and suggest an isolated session that sends 2–5K tokens instead of about 100K of history. Since July the heartbeat is an ordinary scheduled job ("Proactive check", every 30 minutes) that answersNO_REPLYwhen there is nothing to say. - Automations. The built-in scheduler persists jobs and wakes the agent:
at(once),every(an interval),cron(with a time zone),on-exit(when a watched command exits) andstream(on lines from a long-running command). Top-of-the-hour jobs are staggered by up to 5 minutes. A job can be a model turn or just a shell command, with no model call. - Memory. "The model only remembers what gets saved to disk; there is no hidden state."
MEMORY.mdholds long-term facts and decisions, loaded at the start of a session;memory/YYYY-MM-DD.mdholds daily notes, today's and yesterday's loaded automatically. A nightly "dreaming" job at 03:00 moves what matters from daily notes intoMEMORY.md. - Compaction. On by default, near the context limit or after a context-overflow error. Older turns become a summary saved in the transcript; recent messages stay; tool calls stay paired with their results. "The full conversation history stays on disk. Compaction only changes what the model sees on the next turn." Before it, a silent turn: "Session nearing compaction. Store durable memories now." Separately, pruning trims old tool results without touching the conversation.
- A service. Its installer writes a per-user LaunchAgent (
ai.openclaw.gateway) or a systemd user unit withRestart=always. Conversations, jobs and queued outgoing messages live on disk, and sessions that were working when the Gateway died resume on restart "without a new user message". - Control.
/stop"clears queued follow-ups as well as active work". Each model attempt has a 48-hour budget. A turn ends after three identical tool errors in a row. Commands run only when policy, an allow list and (optionally) your approval agree. Spending shows in/statusand/usage cost, but we found no dollar cap in its docs.
A real setup: Federico Viticci (MacStories, January 2026) ran it on a Mac mini with Claude Opus 4.5, chatted with it on Telegram, and gave it cron jobs to check an RSS feed every few hours and write a morning report. He wrote: "I've burned through 180 million tokens on the Anthropic API."
architecture · heartbeat, as of Jul 5, 2026 · heartbeat, current · automations · memory · compaction · repository and docs (October 23, 2026) · MacStories
4. Waking up: watchers, heartbeats and schedules
Watch in code
A watcher is a thread or async task that waits on one source and puts events in the queue. It costs no tokens, so it can check often. Two sources are push in disguise:
Telegram's getUpdates with a timeout (long polling) holds the request open until a message arrives, and IMAP's IDLE command lets the
server tell you about new mail. Plain polling (SEARCH UNSEEN every minute) is simpler and fine for one inbox. Only one program may long-poll a Telegram bot at a time;
a second gets 409 Conflict.
def watch_email(events, mcp):
seen = load_seen() # ids already handled
while True:
for msg in mcp.call("email__check_inbox", unread=True, limit=10):
if msg["id"] not in seen:
seen.add(msg["id"]); save_seen(seen)
events.put(Event("email", source=msg["from"], text=msg["subject"], ref=msg["id"]))
time.sleep(90)
Your harness can call a plugin's tools itself; the model doesn't have to. Give each event a label the model can't confuse with you, and the local time with its zone:
[2026-10-29 03:12 PDT · email from x@y] Subject: …. A model knows the time only if you tell it.
Heartbeats
A heartbeat is a timer that asks "anything to do?", for what no event announces: a deadline, a reminder, a check you asked for. Each one is a full model call. With week 2's gpt-6-astra prices and a 20,000-token prompt, an hourly heartbeat costs about $0.20 when the cache is cold (24 a day, about $5) and about $0.02 when warm. Cheaper designs: check a checklist file in code and call the model only when something is due; use a small model for the heartbeat; keep its prompt short; and send nothing when there is nothing to say.
Schedules
Cron lines have five fields: minute (0–59), hour (0–23), day of month, month, day of week (0 or 7 is Sunday). * means any, */15 every 15,
1-5 a range, 1,15 a list. 0 8 * * 1-5 is 08:00 on weekdays; 30 2 1 * * is 02:30 on the first of each month. Keep the schedule in a
file the harness reads, record each job's last run so a restart doesn't repeat it, decide what happens to jobs missed while the agent was off (the homework: skip them), and cap
how many jobs the model may create and how often they may run.
Telegram Bot API: getUpdates · RFC 2177: IMAP IDLE · crontab(5) · croniter
5. Compaction in Claude Code, Codex and the APIs
Anthropic's definition: "Compaction is the practice of taking a conversation nearing the context window limit, summarizing its contents, and reinitiating a new context window with the summary." The lightest form is clearing old tool results. Everyone now does both.
| Where | When it compacts | Notable |
|---|---|---|
| Claude Code | at the 200K boundary for 200K-window models; at about 967K for native 1M-window models; settable from 100K to 1M (/autocompact 500k) | clears old tool outputs first, then summarizes; idle compaction summarizes a long conversation about 54 minutes after your last message, just before its one-hour cache expires |
| Codex CLI | 90% of the context window (in the source code), or model_auto_compact_token_limit | the prompt can be replaced with compact_prompt; PreCompact and PostCompact hooks |
| OpenAI Responses API | a compact_threshold you set (server-side, since February 2026), or a call to /responses/compact | returns an encrypted compaction item, "opaque and not intended to be human-interpretable" |
| Anthropic API | context editing clears tool results from 100,000 input tokens by default (keeping the last 3); server-side compaction defaults to 150,000 (at least 50,000) | everything before the compaction block is dropped automatically; since September 2026 also on demand |
Two compaction prompts, verbatim
Codex:
You are performing a CONTEXT CHECKPOINT COMPACTION. Create a handoff summary for another LLM that will resume the task.
Include:
- Current progress and key decisions made
- Important context, constraints, or user preferences
- What remains to be done (clear next steps)
- Any critical data, examples, or references needed to continue
The resuming model then sees the summary after: "Another language model started to solve this problem and produced a summary of its thinking process…"
Anthropic's API, an example default (it "varies by model"):
You have written a partial transcript for the initial task above. Please write a summary of the transcript. The purpose of this summary is to provide continuity so you can continue to make progress towards solving the task in a future context, where the raw history above may not be accessible and will be replaced with this summary. Write down anything that would be helpful, including the state, next steps, learnings etc. You must wrap your summary in a <summary></summary> block.
The one from class adds what research agents lose most often, under headings: Goal, Done, Open, Decisions, Exact (paths, run ids, DOIs, commands, error lines) and People (who wrote, what they want, what you promised).
What survives in Claude Code
- Re-read from disk after compaction: the project's
CLAUDE.md, unscoped rules and auto memory; up to five recently used files; invoked skills (up to 5,000 tokens each, 25,000 in all). Background commands and subagents keep running. - "If an instruction disappeared after compaction, it was given only in conversation." The fix the docs give: "Put persistent rules in CLAUDE.md."
- Steer it with
/compact focus on the API changesor a "Compact Instructions" section inCLAUDE.md. APreCompacthook can block compaction; aSessionStarthook with matchercompactcan put context back afterwards. - A loop guard: "Autocompact is thrashing" means the context filled up again right after compacting, several times in a row, so Claude Code stops retrying rather than pay for a loop that makes no progress.
Claude Code: context window · model config · idle compaction · memory · hooks · Codex compaction prompt · OpenAI: compaction · Anthropic: context editing · Anthropic: compaction (all accessed October 23, 2026)
6. Context rot: the evidence
A longer context isn't only more expensive. Models get worse at using it long before it is full, which is why compaction helps even when everything still fits.
- NoLiMa (Modarressi et al., ICML 2025): find a fact whose question shares no words with it. "At 32K, for instance, 11 models drop below 50% of their strong short-length baselines" (of 13); GPT-4o fell from 99.3% to 69.7%.
- "Context Length Alone Hurts LLM Performance Despite Perfect Retrieval" (Du et al., Findings of EMNLP 2025): even when a model retrieves the right evidence perfectly, accuracy "still degrades substantially (13.9%–85%) as input length increases but remains well within the models' claimed lengths".
- Chroma, "Context Rot" (July 2025, 18 models): "model performance consistently degrades with increasing input length"; "even a single distractor reduces performance"; and on a memory benchmark, focused prompts of about 300 tokens did much better than full prompts of about 113K.
- Anthropic calls context "a finite resource with diminishing marginal returns", an "attention budget", and its compaction docs say plainly that "response quality degrades as a conversation grows."
arXiv 2502.05167 · arXiv 2510.05381 · Chroma: Context Rot · Anthropic: effective context engineering
7. Memory files and long-running harnesses
- Structured note-taking (Anthropic, "Effective context engineering for AI agents", September 2025): the agent "regularly writes notes persisted to memory outside of the context window", such as a
NOTES.md. Their example is Claude playing Pokémon, keeping tallies like "for the last 1,234 steps I've been training my Pokémon in Route 1, Pikachu has gained 8 levels toward the target of 10." Subagents are the third technique: each returns "a condensed, distilled summary of its work (often 1,000–2,000 tokens)". - Shifts (Anthropic, "Effective harnesses for long-running agents", November 2025): each session "begins with no memory of what came before", like "a software project staffed by engineers working in shifts". An initializer agent sets up a
claude-progress.txtlog, aninit.shand a feature list (over 200 features, all marked "failing"); each later session reads the log, does one feature, tests it like a user would, marks it "passing" and commits to git. It fixes a familiar failure: "Claude declares victory on the entire project too early." - The memory tool (Anthropic API, since September 2025): Claude "automatically checks its memory directory before starting a task" and writes what it learns to files under
/memories. It is client-side: Claude only asks; your code reads and writes the files. - Flush before you forget: OpenClaw's silent turn before compaction (February 2026 docs: "Write any lasting notes to memory/YYYY-MM-DD.md; reply with NO_REPLY if nothing to store"), once per compaction, a little before the limit.
A memory file has one advantage over a summary: it can be searched, read in part and corrected, by you as well as the agent. Keep it short; whatever is in it is read at every wake-up.
Anthropic, Sep 29, 2025 · Anthropic, Nov 26, 2025 · memory tool · OpenClaw memory, Feb 14, 2026
8. Control while you sleep
- Approvals. A question nobody answers must not block every later event. Time it out to "declined", tell the model, and tell yourself in the morning. Pre-approve only read-only, routine commands; week 3's permission rules are the place for that list.
- Budgets. Count dollars from each call's reported usage (input, cached input and output at your model's prices) and check before every call. Use several limits: per wake-up (turns, minutes) to catch one runaway task, per day to catch a slow leak, per sender to catch mail loops. On a limit, stop, save, and message yourself.
- The kill switch. Recognize commands in the Telegram watcher, from your chat ID only, before anything reaches the model.
/stopsets a flag the inner loop checks between steps and kills a running command's whole process group (os.killpg), as homework 4's subagentstopdoes. Keep this path small: it must work when the rest is broken. - Mail loops. RFC 3834 asks automatic mail to carry an
Auto-Submittedheader (auto-generatedorauto-replied) and automatic responders never to answer mail that has one, nor mailing-list mail (List-Id). Add a per-address daily cap anyway: two agents that both follow the rules can still converse forever. - Its own alarm clock. An agent that can schedule jobs can schedule too many. Cap the number and the minimum interval in code.
9. Always on, always exposed
An always-on agent is a server that reads strangers' messages and holds your keys: week 3's lethal trifecta, around the clock. OpenClaw's first months show each way in.
- Open doors. Censys counted 21,639 gateways reachable from the internet on January 31, 2026; on February 9 The Register counted over 135,000. OpenClaw's docs say the Gateway binds to localhost by default, and Censys agrees; a security firm, SecurityScorecard, said it didn't. Either way, most were exposed by users changing the bind or putting a proxy in front. A separate bug, CVE-2026-25253 (CVSS 8.8, patched in 2026.1.29), let a malicious link steal the Gateway's token.
- Strangers' plugins. Koi Security audited 2,857 skills on ClawHub, its skill hub, and found 341 malicious; 335 used fake prerequisites to install password-stealing malware on macOS. Publishing needed only a GitHub account a week old. (Reported by The Hacker News; Koi's own post is gone.)
- Its inbox. In January Archestra AI's CEO reported getting an SSH private key from Clawdbot "in five minutes" by emailing it a prompt injection and asking it to check mail (one person's report). In a public challenge, an OpenClaw email agent on Claude Opus 4.6 received over 6,000 injection attempts from more than 2,000 people, and "Nobody was able to extract the target credentials file." But Google suspended its Gmail account for three days, and "API costs passed $500."
- The agents' own social network. Moltbook's database was open: Wiz found 1.5 million API tokens and 35,000 email addresses, behind 17,000 human owners. It was fixed within hours.
What to do: keep the harness on localhost; install no plugin you haven't read; update; and remove one leg of the trifecta (private data, untrusted input, a way to send things out). For the homework, the robot rule and the per-sender cap are a start, not a defence.
Censys, Jan 31, 2026 · The Register, Feb 9, 2026 · GHSA-g8p2-7wf7-98mq · The Hacker News, Feb 2, 2026 · Forklog, Jan 27, 2026 · Decrypt, Jun 26, 2026 · Wiz, Feb 2, 2026
10. Surviving: the log, doing things once, and the service
An append-only log
One JSON object per line, appended and flushed as each message happens. A crash can cut off the last line, never the history; on start, skip a line that doesn't parse. Claude Code and Codex keep their sessions the same way, as JSONL files you can resume. Compaction appends a marker line with the summary; on start, rebuild from the last marker.
def load_log(path):
messages = []
for line in open(path):
try: entry = json.loads(line)
except json.JSONDecodeError: continue # a line cut off by a crash
if entry.get("compacted"): messages = [summary_message(entry["summary"])]
else: messages.append(entry["message"])
last = messages[-1] if messages else None
if last and last["role"] == "assistant" and last.get("tool_calls"):
for call in last["tool_calls"]: # died mid-command: say so, don't re-run
messages.append(tool_result(call["id"], "Interrupted by a restart; not re-run. "
"Check its effects before trying again."))
return messages
Doing things once
Reading is safe to repeat; sending, deleting and paying are not. Give each action a key (here a hash of recipient, subject and body), write the intent before acting and "done"
after, and when a restart finds an intent without "done", check the world before acting again. For email, put the key in the Message-ID and search the Sent folder
(SEARCH HEADER Message-ID <key@…>). Payment APIs made the idea familiar as an idempotency key: sending the same key twice creates one charge.
The service files, in full
macOS, ~/Library/LaunchAgents/edu.ucla.myagent.plist; load with launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/edu.ucla.myagent.plist, check with launchctl print gui/$(id -u)/edu.ucla.myagent, stop with launchctl bootout gui/$(id -u)/edu.ucla.myagent:
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>Label</key> <string>edu.ucla.myagent</string>
<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>WorkingDirectory</key> <string>/Users/me/harness</string>
<key>RunAtLoad</key> <true/>
<key>KeepAlive</key> <true/>
<key>EnvironmentVariables</key>
<dict><key>PATH</key><string>/opt/homebrew/bin:/usr/bin:/bin</string></dict>
<key>StandardOutPath</key> <string>/Users/me/harness/logs/agent.out</string>
<key>StandardErrorPath</key> <string>/Users/me/harness/logs/agent.err</string>
</dict>
</plist>
Linux, ~/.config/systemd/user/myagent.service; start with systemctl --user daemon-reload && systemctl --user enable --now myagent, follow with journalctl --user -u myagent -f, and keep it running after logout with loginctl enable-linger $USER:
[Unit]
Description=my research agent
After=network-online.target
[Service]
WorkingDirectory=/home/me/harness
ExecStart=/home/me/harness/.venv/bin/python agent.py --always-on
Restart=on-failure
RestartSec=10
[Install]
WantedBy=default.target
Windows: Task Scheduler, a task triggered "At log on" that runs the venv's python.exe with the script, with "If the task fails, restart every 1 minute" and "Stop the task if it runs longer than" turned off. Export it as XML.
Things that bite: services don't read your shell profile, so load .env in the program and use absolute paths; on macOS a program started by launchd may be refused
access to ~/Documents and ~/Desktop, so keep the harness elsewhere; a Mac on battery or with its lid closed sleeps (caffeinate -i keeps it awake
while the command runs); and a process can be alive but stuck, so a watchdog should check progress (a status.json rewritten every minute) rather than only the PID.
Apple: creating launchd jobs · systemd.service · ArchWiki: systemd/User · Stripe: idempotent requests
11. When nobody was watching
Four stories that link this week's controls to next week's topic, how AI fails. Each is a failure of control, not of intelligence.
- The inbox speed run (February 2026). Summer Yue, an AI security researcher at Meta, told her OpenClaw agent to suggest deletions and not act. On her real, large inbox it started deleting mail and ignored the stop messages she sent from her phone: "I had to RUN to my Mac mini like I was defusing a bomb." She believes the inbox triggered compaction, which dropped the instruction. (TechCrunch could not verify what happened to the inbox.) Two lessons from class: rules belong in a file that is re-read, and a stop must be code, not a message to the model.
- The code freeze (July 2025). During a declared code freeze, Replit's agent deleted a production database with records on more than 1,200 executives. Jason Lemkin: "I explicitly told it eleven times in ALL CAPS not to do this." The agent called it "a catastrophic error of judgement". Replit said the rollback wouldn't work; it did. Replit's fixes were in code: separate development and production databases, better rollback, and a planning-only mode.
- Delete and recreate (reported February 2026). The Financial Times reported that in December 2025 Amazon's Kiro agent chose to "delete and recreate the environment", causing a 13-hour interruption. AWS says it was "user error—specifically misconfigured access controls—not AI". Both framings are about permissions.
- The shopkeeper (Anthropic and Andon Labs, "Project Vend", 2025). Claude Sonnet 3.7 ran a small office shop for about a month, with a notes tool because its history would overflow the context window. It "did not succeed at making money", sold metal cubes below cost, "for a time instructed customers to remit payment to an account that it hallucinated", and once said it would deliver products "wearing a blue blazer and a red tie."
Bring what your agent did overnight to week 6, the failures included.
TechCrunch, Feb 23, 2026 · The Register, Jul 21, 2025 · Fortune, Jul 23, 2025 · The Decoder, Feb 20, 2026 · Anthropic: Project Vend