text_description.md

Kinloop — text description

(This is the submission text description required by the hackathon rules: what it does, who it's for, and how it works.)

What it does

Kinloop is an autonomous agent that manages the day-to-day coordination burden of family caregiving. Every day, unattended, it:

1. Checks every tracked medication for a shared care recipient and flags — or auto-drafts a refill request for — anything overdue or running low.
2. Scans the family's shared appointment calendar for scheduling conflicts: a family member double-booked as a driver, or an appointment with nobody assigned to take the patient. It proposes a fix using who else is free.
3. Tracks recurring paperwork deadlines (insurance recertification, benefit renewals) and flags what's missing before it's too late.
4. Decides — deliberately, as its own dedicated step — which of the above is routine (silently logged) versus a genuine decision that needs a human, and only notifies a family member for the latter.

A good day for Kinloop is a quiet one: everything routine gets handled or logged, and the family only hears from it when something actually needs their judgment.

Who it's for

Families sharing caregiving duties for an aging parent or a relative with a chronic condition — typically 2–5 people (adult children, a spouse, in-laws) spread across different households, which is exactly the situation where information gets lost between people. In the US alone, roughly 1 in 5 adults is currently providing unpaid care to a family member, most of it exactly this kind of fragmented, repetitive coordination work: medications, appointments, and paperwork.

Why it matters

Every existing tool in this space — shared calendars, caregiving apps like CareZone or Lotsa Helping Hands — is a passive tracker. A human still has to open it, read it, and reconcile what's happening across three different domains (meds, appointments, paperwork) and three different people's schedules. None of them act. That's the actual gap: this is a background-agent problem, not an app problem, and it hasn't been built as one. Kinloop is built to run unattended and only interrupt a human for genuine decisions — which is also the literal design brief of this hackathon's "Everyday Agents" track.

How it works

Kinloop is built with the Strands Agents SDK using the Agents-as-Tools multi-agent pattern: a Supervisor agent delegates to four specialist agents (MedAgent, ApptAgent, PaperworkAgent, NotifierAgent), each with its own system prompt and tools, called as tools themselves from the Supervisor. NotifierAgent is a deliberately separate agent — a dedicated escalation gatekeeper — rather than folding that decision into each specialist, so "should a human be interrupted" is made once, consistently, by comparing findings across all three domains together.

Deterministic logic (is a refill overdue, do two appointments conflict, is a deadline urgent) is implemented as plain, unit-tested Python functions that the Strands @tool decorator exposes to the agent loop — the LLM reasons about what to do with that information (draft a message, decide who to escalate to, phrase a notification), not about arithmetic on dates.

Runtime: AWS Lambda, invoked once daily by EventBridge Scheduler, backed by DynamoDB and Amazon Bedrock (Amazon Nova Lite, on-demand). A Strands BeforeToolCallEvent hook enforces a hard ceiling on how many sub-agent calls the Supervisor can make per run, as a cost safety measure independent of prompting.

Why not Bedrock AgentCore Runtime for the always-on path: AgentCore's Runtime, Memory, Gateway, and Observability components bill independently and have essentially no free tier beyond a new account's general credit — and Observability specifically is uncapped, with idle-session memory accrual as a documented common source of surprise bills. For a background agent that runs for seconds once a day, Lambda + EventBridge is both cheaper and safer: near-permanent AWS free tier, and zero cost while idle. Kinloop still demonstrates AgentCore Runtime deployment (deploy/agentcore/) as a one-time, intentionally torn-down deployment, showing the same agent code running on AgentCore without being the production path.