Agentry

Agents for Humans Hackathon · AWS × Strands · Everyday Agents

The pantry restocks
itself.

Hand it a goal. It plans the order, drives a real storefront, pays from the platform wallet, and pings you only when a person needs to decide.

View the repoHow it works
$python main.py --goal "restock the pantry"

Product

Everything a pantry run needs, done for you

One goal in, one order out. Agentry plans, shops, pays and reports back — you only hear from it when it counts.

  • 01

    Plans the order

    Hand it a goal in plain language — “restock the pantry.” It turns that into a real cart: quantities, substitutes, a budget it won’t blow past.

  • 02

    Shops for real

    No sandbox, no mock checkout. It opens an actual quick-commerce storefront and drives it end to end, DOM changes and all.

  • 03

    Pays like you would

    Checkout settles from the storefront’s own wallet balance. No cards handed around — just the money that’s already there.

  • 04

    Speaks up only when it matters

    One Telegram thread, saved for the moments a person actually has to decide — an out-of-stock staple, a price past the ceiling.

Why it’s different

Built to stay out of your way

Most “assistants” ask you to confirm every step. Agentry inverts that: it acts, and saves the interruption for the one moment it matters.

  • Real storefront, not a mock

    It automates an actual quick-commerce site — selectors, stock, prices, all of it live.

  • Platform-wallet checkout

    Pays from the storefront’s native balance. No card entry needed.

  • A budget for interruptions

    One Telegram thread, spent only on genuine decisions — never step-by-step narration.

  • Genuine Strands

    Agent, tools, model providers and hooks — a real orchestration layer, not a wrapper over an API call.

03 · Architecture

One agent, a handful of tools

A goal enters as plain text. The Strands agent plans it, then works a real storefront through its tools — surfacing on Telegram only when a person has to decide.

Agent workflow
Stretch · Bedrock AgentCore Runtime
Intent

Goal

  • Plain-language brief — “restock the pantry”
  • Resolved into a concrete cart, budget and substitutes
Step 01
Orchestration

Strands Agent

Plans the run, calls tools in a loop, and reflects on each result — with a local knowledge graph for past orders.

Agent(
  model=GeminiModel(...),
  tools=[browser, telegram, cart],
)
Browser

Playwright

Stealth Chromium on a live quick-commerce storefront — search, compare, add to cart, ride out DOM changes.

Step 02
Human-in-the-loop

Telegram

One thread, used only when a person must decide — a missing staple, a price past the cap.

Fulfilment

Storefront & Wallet

  • Cart assembled on the real Zepto storefront
  • Checkout paid from platform cash — no card needed
  • Order lands; a Telegram receipt closes the loop
Step 03 · done
Memory

Knowledge graph

Past orders — items, brands, prices, accepted substitutes — read and written on every run.

Strands Agents SDK · Gemini · Python · Playwright · Telegram Bot API · AWS Bedrock AgentCore

Under the hood

What it’s made of

No proprietary glue — a small stack of tools doing exactly what they’re good at.

  • Strands Agents SDK

    Agent, @tool, model providers and hooks — the orchestration layer, rebuilt from scratch.

  • Gemini

    Planning and reasoning via Strands’ GeminiModel provider.

  • Playwright

    Stealth-configured Chromium driving the live storefront end to end.

  • Telegram Bot API

    The single human-in-the-loop channel — used sparingly, on purpose.

  • AWS Bedrock AgentCore

    Stretch target for hosted deployment of the Strands agent.

Get started

Stop restocking it yourself.

Clone the repo, point it at a storefront login, and give it a goal. The next pantry run is on it.

View the repoOpen the console
$python main.py --goal "restock the pantry"