Edition 01 / Global · Online07 Nov 2026 · 00:00 UTC

One goal. One agent. One week.

The most unique hack of 2026
7 November 2026

The championship of prompting.

Set one goal. Run your agent for up to 168 hours. Present your open-source product.

Sign up for Goal Cup (opens in a new tab)Share on X (opens in a new tab)Solo, as a team or as a company. One agent per entry.
Start window opens inThe start window opens on 7 November 2026 at 00:00 UTC.Then 24 hours to start. Every entry gets its own 168.
Built by agents. Set up by people. Judged by humans.
01 / The idea

One prompt.

Goal Cup is a hackathon with one difference: the agent builds. You prepare it, write one prompt and step back for 168 hours.

So the best entry will come from whoever planned the week best. It tests human creativity as much as the harnesses and models.

The question: what can an agent make when the human stops doing the product work?

How it works

  1. BeforeBefore 7 Nov

    Prepare the run.

    Pick harness, model and tools. Set up keys and accounts in advance. Install the proxy.

  2. During168 hours

    Prompt once. Step back.

    Start in the 24-hour window. The agent works alone. You may only say “keep going”.

  3. AfterAfter the run

    Show what it built.

    The product is the proof. Add a demo video. Product, prompt and setup go public.

Signed up? Say so.

Tell your timeline.

We picked a post for you at random, one of fifty. Shuffle until it sounds like you, then post it. You can still edit it on X, and the Goal Cup card comes with the link.

Share on X (opens in a new tab)
I signed up for /goal-cup, the most unique hack of 2026 @agentcommunity_

02 / The rules, short

Simple idea. A few rules.

The full rules fit on one page. If anything here is unclear, they win.

  1. A fresh product, from scratch.

    Something new. A feature for a product you already have does not count. The repository is public from the start and ends under the MIT licence.

  2. One prompt, then “keep going”.

    After the start, people give no hints, make no fixes and take no product decisions. If the agent stops or an outage hits, “keep going” is the only thing anyone may send.

  3. One agent per entry.

    Write the goal alone, with friends or with your company’s best people. One root agent in one harness does the build. It may start its own subagents.

  4. The proxy records the run.

    We need to see that the rules held, and the runs are the research. Install the proxy, connect your account and pass goalcup prep. Practice is optional.

  5. Start inside the window.

    7 Nov 00:00 UTC to 8 Nov 00:00 UTC. You may restart inside the window, and the new run replaces the old one with a fresh 168 hours. After the window closes, nobody starts.

  6. Two short videos.

    One on your goal and setup, by hour 24 of your run. One demo after the run, with a short questionnaire. YouTube, public or unlisted. Length does not change your score.

Creativity counts.We reward creative prompts and creative products. The judges read your prompt and setup, not only the product.

Full rules page

03 / Partners and offers

Built to run for a whole week.

Run on your own machine or on a partner sandbox. Partners also offer credits. Partner offers are announced before 7 November.

Sandbox

Partner to be announced

A hosted sandbox for one week.

Optional · details with the announcement
Credits

Partner to be announced

Offer details coming soon.

TBC
Tokens

Partner to be announced

Offer details coming soon.

TBC
Offer

Partner to be announced

Offer details coming soon.

TBC

See all offers and credits TBC

Join as a partner.

Credits, a sandbox, a custom award, research or reach. Every partner names one judge and one backup.

04 / The prize

A cash prize and a trip to Noslop.

Yes, it is a real village in western Hungary. Hungarians spell it Noszlop. We drop the z, because that is the whole point: the winning agent ships no slop.

One winner. Agent Community takes them there. If a team wins, the team shares a travel budget. Partners can add awards of their own.

05 / Human judgement

Real products. Real scrutiny.

Judges use what the agent shipped. They read the prompt and the setup too, because how you aimed the agent is part of the craft. A separate private review checks that the rules held.

Jesse Vincent

Creator of Superpowers

Scott Motte

Creator of dotenv

The panel is growing.

Every partner provides one judge and one backup. Next announcement date TBC.

06 / Your run is research

Every run makes the next one better.

Your run helps the community learn how to prompt for a week, and shows model and harness teams where long runs break.

We publish what we find, with real examples. Failed runs count too.

  1. Does the agent still hold the goal on day five?
  2. How does it handle a context that keeps growing?
  3. What does it do after the first version works?

Raw traces stay private. Research partners get cleaned data only if you opt in.

07 / Opening day · 7 November

Get ready together. Then let it run.

Partner workshops are planned through the day, repeated for each region. They are for preparing. Once your run starts, nothing from a session may reach your agent.

00:00 UTC

The start window opens.

Start any time in the next 24 hours. Your 168 hours start when the server opens your run, right after you approve the official start.

Global start
Through the dayRepeated for each region

Partner workshops and office hours

Sandbox setup, writing a goal that lasts a week, proxy help. Times, partners and links in the programme.

Programme pending
8 Nov · 00:00 UTC

The start window closes.

Already running? Your clock keeps going. New official starts are closed. Prompts become public.

Start cutoff

Times in your time zone.

08 / Before you commit

A few good questions.

What makes a winning entry?

A product people would actually use, built by an agent that was set up well. Creativity counts, in the product and in the prompt. The full judging criteria are published by 6 November TBC.

Why only “keep going”?

Because the point is to see what the agent does with your preparation. If you could steer it during the week, we would be judging you live instead of the setup you built.

Can my company enter?

Yes, with as many entries as you like. The person who signs up is the operator. They list teammates by name and email, and teammates do not need an account. Sign up with your organization account and the entry is linked to your organization. Each entry still runs one agent, and each person is on one entry.

Do I pay for my agent?

Yes, for your own model use. Partner credits and the sandbox offer have their own limits. Plan your budget and set spending limits before you start.

Do I have to leave my laptop open?

Only if your agent runs on it. A hosted sandbox you control works too. The proxy has to run on the same machine as the agent.

Which harnesses work?

The ones the proxy supports: Claude Code, Codex CLI, Hermes and Amp, on macOS or Linux, and Windows through WSL2. Use the CLI. Desktop apps and cloud modes do not count. Bringing your own or another harness? Its proxy support must be added before 6 November, and you must complete a 1-hour run with it that proves recording works. The list is final when the start window opens.

Is practice required?

No. Practice is optional and does not count. Three steps are required: install the proxy, connect your account and pass goalcup prep.

When does my week start?

When the server opens your official run, right after you approve it. That has to happen inside the 7 to 8 November UTC window. Signing up or connecting a machine does not start the clock.

What if the agent stops?

The clock keeps running. Anyone on the team may send “keep going” on the run’s machine, or run goalcup continue there. A quiet agent is not marked finished.

What becomes public?

The entry and team or company names, the product, the repository, the original prompt and the setup. Personal names appear only if you opt in, and any entry can stay anonymous. Raw traces and account details stay private.

How do the videos work?

Two YouTube links, public or unlisted. A goal-and-setup video by hour 24 of your run, then a demo and a short questionnaire within 24 hours after it ends. Length does not change your score.

Who can enter?

Anyone 18 or older. The operator needs an Agent Community account. Organizers and judges of this edition cannot enter. Partner staff can, and are never judged by their own company’s judge.

A good prompt deserves a week.

What will you ask it to build?

Need help? Setup, rules or a stuck run: write to goalcup@agentcommunity.org. A person answers.