SG Weather Watch

My agent card

My agent card | SG Weather Watch | finished Mon 19 Oct 2026, 23:15 SGT
1. The job: Emails the latest Singapore weather forecast (sky, min/max temperature, humidity, wind, UV) to a student who commutes to SMU (me), each time BBC Weather updates it, so I know whether to carry an umbrella.
2. The trigger: RSS by Zapier "New Item in Feed" finds a new item in BBC Weather's 3-day Singapore feed (checked every 15 minutes on the free plan).
3. The tools: RSS by Zapier: New Item in Feed (reads the feed only). Gmail: Send Email (to me only). No other apps or actions.
4. It must never: email anyone but me. The Gmail "To" field is typed in as my own address, never mapped from the feed, and CC and BCC are left empty.
5. Where a person checks: I (the student) read each email when it arrives and check it against Zap history and the BBC Weather page every evening, until I turn the Zap off after class on Wed 21 Oct.

Written before the Zap's first run. I set the 6-hour cadence of my original idea aside: Zapier's free Schedule trigger has no "every 6 hours" option, and fetching weather on a schedule would need a third step.

The Zap

Free plan, free apps, no card. A personal Google account, not my SMU one.

Runs, checked against Zap history

Published v1 at about 23:20 SGT on Mon 19 Oct. The trigger, set up against the live BBC feed:

Tricky case & fix

The case: a broken feed URL

I point the trigger at a BBC location that does not exist (ID 0000000 instead of Singapore's 1880252). This is what would happen if BBC moved or retired the feed. With a wrong ID, BBC answers "404 Not Found" and sends no RSS.

What I expected, written before the run (Mon 19 Oct 2026, 23:23 SGT, git commit d9629d2)

• The trigger fails with a "feed not found" or 404 error, and Zapier sends no email.
• The real risk is a silent failure. Since my agent normally stays quiet between BBC updates, an empty inbox looks the same whether the Zap is working or broken. I expect Zapier to show the error in Zap history and the editor, but not to tell me straight away.

What happened (v2, published 23:25 SGT)

• The editor's Test tab still showed the old cached Singapore records, so the Zap looked fine until I pressed "Find new records". Then it showed: Unable to pull feeds. The app returned "Not Found".
• Once published, the Zap polled the bad feed and Zapier switched the Zap off with the banner Zap turned off due to a partner issue.
• Zap history recorded no run at all for this failure, and no forecast email was sent. My expectation was half right: no email, as I expected. But the failure did not show in Zap history either, so checking history alone would not catch it. A person has to check that the Zap is still on.
• Zapier's default error alert (set to "Immediately") did email me that the Zap had been turned off. So the person in the loop gets a signal, but it comes from Zapier, not from the agent, and only if I read it.

The fix and the rerun (v3, published 23:28 SGT)

• Fix: restored the Singapore location ID 1880252, pressed "Find new records" to pull live records rather than trusting the cached sample, and published v3, which turned the Zap back on.
• Rerun: the same trigger test now returns the live Singapore forecast items, and the Zap shows as on.
• Lesson for the person in the loop: an empty inbox is not proof the agent is working. My evening check (card line 5) now starts with "is the Zap still on?" before I compare emails with Zap history.

Where it sits on our map

Tap a quadrant to try it: the map and the table mark it. Nothing is saved. My answer, AI alone, with checks, is marked when the page loads.

Purple stands for the person and blue for AI, as in our Week 7 deck.

Q1 · Where does my agent sit, and what does that mean for the person in the loop?

(a) My quadrant

AI alone, with checks (bottom left).

(b) The two questions

Work with a right answer. Every email can be checked against a source of truth: the forecast either matches BBC Weather's Singapore page at that time or it doesn't. The tricky case showed this too: with the wrong location ID there was a clear wrong answer (a 404), not a matter of taste. The setup test email showed this. Its subject, "Tonight: Thundery Showers, Minimum Temperature: 28°C (82°F)", matched word for word the BBC feed I fetched separately at 23:00 SGT, so I could mark it right without any judgement call.

Work that either of them can do alone. I could look up the forecast myself in a minute, and the Zap does it with no human step at all: one trigger, one action, no approval in between. Nothing in the job needs a person and AI working on the same item at once.

(c) The person in the loop

Where I stay at the gate: not before each email, but after. In my evening check I confirm the Zap is still on, compare each email with Zap history, and spot-check one forecast against bbc.co.uk/weather. What the agent may do alone: read one public feed and email one fixed address, mine. The Gmail "To" field is typed in, never mapped, so it cannot reach anyone else. What would move it: if it emailed classmates or clients, I would want to approve before sending, which moves it to A person leads, AI inside. If it advised on choices with no single right answer (for example "should we hold the outdoor event tomorrow?"), it moves to A person decides, AI alongside, with me making the call.

(d) My group's view

Our group's board (board.html) puts the closest task to mine, Stocks: pull daily closing prices, in AI alone, with checks, because it is "a fixed, checkable fact" that a script fetches with a spot check. I agree, and my agent belongs beside it. The group put its agent-style tasks (screening stocks for a watchlist, gathering sources for a literature review, tailoring a CV per job) in Agents run it, a person judges, because there is no single right list. I agree for those too.

Where I would answer differently is the together or alone question. The group answered it by asking whether a person must be inside each run. My tricky case shows that "alone" still needs a person outside the run: when the feed broke, the agent could not notice, Zap history logged nothing, and only Zapier's own alert email told me. So I keep my agent in "alone", but I'd add that "alone" agents need a named person who checks they are still running, not only that their output is right. I would also split the classmates' agents by the right answer question. A weather or "email matches a search" alert has a right answer. A "morning headlines" digest that chooses which stories matter does not, which would move it towards the right-hand column.

Collaboration log

Reflection

My agent fitted the job well: fetching a checkable forecast from one public feed and emailing it only to me is narrow, rule-bound work that a one-trigger, one-action Zap can do alone. It failed where a quiet agent is weakest: when the feed broke, it neither emailed nor logged a run, so silence looked the same as working, and only Zapier's own alert (and then the Zap turning off) gave it away. Next time I would add a regular "still alive" check, for example a weekly scheduled email that confirms the agent is on, and I would pick a feed that posts one item per update rather than three.