
Agents for Humans: What I Learned Building an Agent That Knows When to Shut Up
An autonomous agent that runs a family's caregiving coordination — medications, appointments, paperwork — and only interrupts you when there's a real decision to make.
I want to start with the part of this story that has nothing to do with AWS, because it's the part that actually made me build this.
A few years ago I watched a friend of mine coordinate care for her mother almost entirely through group texts. Three siblings, one shared Google Calendar nobody trusted, and a running joke that wasn't really a joke — "did anyone actually call the pharmacy, or are we all just assuming someone did?" Nobody was lazy. Nobody was careless. There were just three people, in three different apartments, each holding a third of the picture, and no single person whose job it was to hold all of it.
That's the thing that stuck with me when I read the brief for the Agents for Humans hackathon. Everyone talks about AI agents that write your emails or summarize your meetings. Almost nobody talks about the unglamorous, exhausting, entirely un-automated work of just... keeping track of another human being's life when that job is split across a family instead of owned by one person. So I decided to build something for that specific gap, and I called it Kinloop.
The idea, in one sentence
Kinloop is an agent that runs once a day, quietly, for a family sharing caregiving duties — checking medications, appointments, and paperwork deadlines — and it only says something to a human when there's an actual decision to make.
That last part is the whole point, and it took me longer than I'd like to admit to realize it was the point. My first instinct, like I think most people's would be, was to build something that told you everything. Refill in 5 days! Appointment tomorrow! Deadline in 3 weeks! And about two days into building that version, I realized I'd just built a more elaborate way to ignore notifications. If an agent pings you for routine stuff as often as it pings you for real stuff, you stop reading it within a week. I've done this to myself with at least four apps already.
So the actual design problem wasn't "how do I track this information." It was "how do I decide what's worth interrupting a person for," and that turned out to be a much more interesting problem to hand to an agent than I expected.
How it's actually built
Kinloop runs on the Strands Agents SDK, and I leaned hard into a pattern called Agents-as-Tools — instead of one big agent trying to reason about medications and appointments and insurance paperwork all at once (and doing a mediocre job at all three), there's a Supervisor agent that delegates to four specialists:
- MedAgent checks refill status and drafts refill requests
- ApptAgent looks for scheduling conflicts — someone double-booked as a driver, or an appointment nobody's covering
- PaperworkAgent tracks recurring deadlines like insurance recertification and flags what's missing
- NotifierAgent — and this is the one I actually care about — takes everything the other three found and decides, deliberately, what's routine and what's a real decision
I split NotifierAgent out on purpose instead of letting each specialist decide for itself whether to escalate. If MedAgent decides on its own whether refills are urgent, and ApptAgent decides on its own whether conflicts are urgent, you get three different, slightly inconsistent definitions of "urgent" scattered across the system. Having one agent whose entire job is the escalation call, looking at everything together, made the whole system behave more like a single thoughtful person and less like three well-meaning strangers all texting you at once. Which, if you think about it, is exactly the problem I was trying to fix in the first place.
Underneath all of that, the actual "is this refill overdue" and "do these two appointments conflict" logic isn't handled by the LLM at all — it's plain Python, fully unit tested, with zero AWS dependency. The model's job is to reason about what to do with that information: how to phrase a refill request, who to escalate a conflict to, how to word a message so it actually gives someone a starting point instead of just anxiety. I didn't plan this split as some clever cost-optimization move at the start. It just felt wrong to spend a model call figuring out whether 3 minus 5 is negative. It turned out to matter a lot more than I expected, for a reason I did not see coming.
The part where AWS almost ate my timeline
Here's the honest version of the story, because I think it's more useful than the polished one.
About halfway through the build, on a brand new AWS account, I hit a wall that had nothing to do with my code: new-account verification. Bedrock calls started failing with an authorization error, and so did spinning up CloudShell. I opened a support case. I got a boilerplate reply. I tried live chat. Days went by. My account was, functionally, in a waiting room, and my hackathon deadline was not waiting with it.
I won't pretend that wasn't stressful. There's a very specific kind of dread that sets in when the thing blocking your project isn't a bug you wrote, it's an account status you can't fix by writing better code. For about a day I genuinely didn't know what I was going to submit.
And then I remembered the thing I'd already built without realizing why I'd built it that way: the deterministic logic underneath every agent didn't need Bedrock at all. So I built a
--dry-run mode that runs that exact same logic — refill checks, conflict detection, deadline urgency, the whole escalation policy — directly, with no model call, and labels every single line of output so nobody mistakes it for the real thing. It's not a workaround I'm proud of hiding. It's right there in the docs, explained plainly: here's what this proves, here's what it doesn't, here's why it exists.What I actually learned from that scramble is that the boring architectural decision I made early — keep the deterministic stuff deterministic, don't route everything through the model just because you can — is what saved the project. Not a clever trick. Just a clean separation of concerns that happened to pay off in a way I didn't plan for.
What I'd tell someone starting this hackathon today
Build the boring parts first, and build them so they don't need the exciting parts to prove they work. Test your logic without your LLM in the loop wherever you honestly can — not because Bedrock will necessarily give you trouble, but because a system where half of it can be verified with zero external dependencies is just a more trustworthy system, full stop. That's true whether or not your AWS account decides to have main character energy for a week.
And build for the moment when your project is quiet, not just the moment when it's impressive. It's easy to design the demo where the agent finds three urgent problems and looks brilliant. It's much more useful — and honestly much harder — to design the version of your agent that finds nothing wrong on a normal Tuesday and has the discipline to just... not say anything. I think that restraint is the actual product, more than any individual feature is.
Where Kinloop goes from here
Sample data was the right call for a three-week build, but the real version of this needs real integrations — pharmacy refill APIs instead of a JSON file, calendar sync instead of a shared fixture, actual document OCR for the paperwork side instead of keyword matching. I'd also like NotifierAgent to learn a specific family's escalation preferences over time instead of applying one fixed policy to every family, because I already know from exactly one conversation with my friend that some families want to be looped in on everything paperwork-related, and others only want to hear from you when it's truly urgent. That's a judgment call, and judgment calls are, it turns out, the thing agents are actually good for — not replacing the group text, but finally being the one person in it who's read the whole thread.
Kinloop was built for the Agents for Humans Hackathon, Everyday Agents track, using the Strands Agents SDK. Code's on GitHub, MIT licensed, if you want to see how any of this actually works under the hood.
#AgentsofFootball
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article