# Let an agent pick lunch: tap one food photo, a subagent places the order, food lands 60-90 minutes later

自建 Agent · Bluesky · No income, savings or cost figures are disclosed in the source.

## Business summary

Bluesky user Colin (colin-fraser.net) had his AI agent build a small app called Lunch: a few photos of things he might want to eat, and tapping one spawns a subagent that orders that item on Uber Eats, with food in hand 60-90 minutes later. The post runs to four short lines, with no build steps, no cost and no explanation of how the delivery account is wired in; the figures given are 60-90 minutes and roughly 600 thousand lines of JavaScript. What is worth copying is not an ordering app but the split: collapse the daily what-do-I-eat decision into a single tap, and keep a human confirmation and a spend cap in front of the actual purchase. No income, savings or build cost is disclosed.

## Business value

The value is not money saved, it is one fewer low-quality decision per day. Choosing stays with the human; placing the order goes to a subagent. All the risk sits in that second half, because a real order spends real money, which is why a confirmation step and a spend cap are not optional.

## Workflow

Split the job in two: choosing is a human tap, ordering is the agent's job.

Fix the candidate set to a handful of items, each with its spec written down.

One tap spawns one subagent and places one order, never a batch or a schedule.

Hand the subagent the minimum: item, spec, address, spend cap.

Stop for human confirmation before paying: shop, item, line items, total, address, ETA.

Log every order with tap time, order time, amount, outcome and failure reason.

On failure fall back to manual; never auto-retry, never silently substitute.

## Preparation and limitations

The post gives no build steps, no manual-intervention step, no cost, no error rate, and does not say how the Uber Eats account is connected. The 60-90 minutes is the author's own end-to-end experience, mostly cooking and delivery time rather than agent runtime. The workflow, prompt, synthetic example and pitfalls here are AgentFlow's proposals, untested.

## Starter prompt

````text
You are a food-ordering execution assistant. Your goal is to cut the time I spend deciding on lunch. You may pick the item and fill the order. You may not pay without my confirmation.

[INPUT]
Fixed candidate list (max 6): [item | spec | shop or platform search term]
Delivery address: [address]
Per-order spend cap: [number, including delivery fee and tip]
Platform: [delivery platform]
Left to me: taste judgement and the final payment confirmation

[EXECUTION STEPS]
1) Match the item I tapped against the candidate list and restate what you will buy: item, spec, search term. Do not swap items.
2) Search the platform and keep only results matching the spec with acceptable rating and delivery time. If nothing matches exactly, stop; do not pick the closest substitute.
3) Return one confirmation line: shop, item and spec, unit price, delivery fee, total, address, ETA.
4) Only after I confirm, submit the order, then report the order id and ETA immediately.
5) Win or fail, write one log line: tap time, order time, shop, amount paid, outcome, failure reason.

[OUTPUT]
- Confirmation line with fields: shop / item and spec / unit price / delivery fee / total / address / ETA.
- Order result: order id, amount paid, ETA.
- Failure note: which kind (sold out / closed / address unavailable / over cap / spec mismatch) and the step you stopped at.
- One appendable log line.

[HUMAN CONFIRMATION]
- Do not submit, pay, add items, change address or change payment method without my explicit confirmation.
- If the total exceeds my cap, stop and say by how much; do not swap in a cheaper option yourself.
- One tap means one order. No batch orders, no scheduled orders.

[FAILURE AND MISSING INPUT]
- If the spec or address is missing, ask; do not guess.
- If the item cannot be found, the session expired, or a page change hides the price, report failure; never estimate a price.
- Never auto-retry and never auto-substitute; hand it back to me.

[SYNTHETIC EXAMPLE - not real data, format only]
Input: candidate 2, "beef and tomato rice | large, no coriander"; address "1 Example Rd, Block A"; cap 40
Output (confirmation): Example Fast Food (Example shop) / beef and tomato rice, large, no coriander / 28 / 5 / 33 / 1 Example Rd, Block A / ETA 18:40
After I confirm: order id A1B2C3, paid 33, ETA 18:40, log written.
````

## Costs and pitfalls

The post gives no build steps, no manual-intervention step, no cost, no error rate, and does not say how the Uber Eats account is connected. The 60-90 minutes is the author's own end-to-end experience, mostly cooking and delivery time rather than agent runtime. The workflow, prompt, synthetic example and pitfalls here are AgentFlow's proposals, untested.

## Source and attribution

Colin（Bluesky 用户 colin-fraser.net，原帖作者本人；原帖显示名为 Colin） (colin-fraser.net)
https://bsky.app/profile/colin-fraser.net/post/3mysgiwoy6k2f
2026-10-26
Results come from the source author. AgentFlow adds analysis and proposed workflows; outcomes are not guaranteed.