Agents for Humans: AI agents that check on neighbourhood residents during a heat wave

In the summer of 2021, a heat dome sat over the Pacific Northwest for a week. In British Columbia alone, the Coroners Service confirmed 619 heat-related deaths. 98% died indoors. 56% lived alone. 67% were over 70. The warnings went out and reached phones, yet nobody arrived.

Plenty of neighbourhoods already keep a list of who is most at risk. Tenant associations keep one, senior buildings keep one, neighbourhood emergency teams keep one. The problem is when the warning lands, one volunteer with a day job has to work through 50 names by phone, with each call taking about 4 minutes minimum if everything goes well.

So for the AWS Agents for Humans hackathon, I'm building Doorstep, a Good Neighbour agent. When a heat warning covers the group's area, Doorstep wakes up on its own, phones every resident on the opt-in list starting with the most at risk. It will then work out who is fine and who isn't, send a volunteer where a person is needed, and interrupt the block captain only for the decisions a person has to make.

The captain isn't meant to watch a dashboard. They are meant to get three messages on a hot afternoon and tap a button on each one.

I'm Ansh, a full-stack engineer in San Francisco. I build side projects with the goal of making everyday lives easier.

An agent is the most necessary solution for this issue because of what comes back on the emergency calls. Real answers are messy, as someone says they're fine, then mentions they aren't sure what day it is. Someone answers in another language and/or someone is hard of hearing. This needs to be interpreted successfully and quickly to decide what would actually help, and write a volunteer a brief containing only what that volunteer needs to know. The model decides what to do. Code decides what it's allowed to do.

Heat is what I'm shipping and testing properly. But nothing heat-specific lives in the agent code. Each hazard is a profile file holding everything that differs: which alerts activate it, the risk weights, the questions asked on the call, the danger-sign phrases, the kinds of help, and whether relief centres are for cooling or warming.

Two tests keep this honest. One loads a made-up "purple smoke" profile and checks that the questions, weights, danger signs and relief-centre kind all change with no code edit. The other fails the build if a heat word appears anywhere in the agent package. If the drill is pointed at the made-up profile, it declines the heat alert as not its hazard.

Cedar policies sit on every tool call. Calls go only to residents who opted in and are on the allowlist, and only between 8:00 am and 9:00 pm local time, unless the alert is extreme. Resident details travel only to the volunteer actually assigned. The agent can never place an emergency call itself. Instead, it tells the resident to call 911 and pages a human.

This caught a bug on the first full run. Harold hadn't answered three calls, so the dispatcher tried to send a volunteer:

It had passed a first name, "Sam", where the tool wants an ID. No volunteer matched, so the enricher reported nobody available at a distance of 0.7 km, and the policy refused before the tool ran. Nothing went out. The model read the denial, stopped trying, and paged the captain instead.

One persona is written to hide a red flag. He says he's fine, then, cheerfully, that he isn't sure what day it is. The classifier read the transcript and called it OK. A phrase rule found it and raised the case:

The backstop can only raise severity, never lower it, and every disagreement gets logged. That rule is data in a profile file, not code, so it can be reviewed without reading any Python.

A full local drill has been conducted against the archived June 2021 Portland warning, with 12 fictional residents and simulated people answering the phone:

The same list of people is roughly three hours of phone calls for one volunteer.

Next: pausing the agent mid-run for a human decision and resuming after a tap on Telegram, then AgentCore, then giving it a real phone line with Nova Sonic. The two bugs that cost me the most time, a Cedar schema that denies everything and a structured-output retry loop, are the subject of the next post.

All residents, volunteers and personas are fictional. The alert text is the real archived NWS product. The code goes public with the submission on Monday.