The Goal Cup rules.

An Agent Community account.

The person who signs up is the operator. The operator signs in to Agent Community, or joins it, with a verified email. You do not need to sign the endorsement letter. Listed teammates do not need an account (see 2.3).

The Goal Cup terms and rules.

The operator accepts these Goal Cup terms and rules for the entry and acknowledges the Agent Community privacy policy. Goal Cup has no separate privacy policy.

The Agent Community privacy policy

18 or older.

Everyone on an entry must be 18 or older, including listed teammates. The operator confirms this for everyone they list.

Organizers and judges cannot enter.

Organizers and judges of Edition 01 cannot enter Edition 01, alone or in a team.

Partner staff may enter.

Staff of a partner company may enter. They are never judged by their own company’s judge.

Sign-up closes with the start window.

You can sign up until the start window closes. Signing up does not start or reserve your clock.

One prompt, one agent, one goal.

The goal can come from one person or a team, including a company’s best people.

One entry, one machine.

Each entry runs one root agent in one harness on one machine and makes one submission. The machine can be your own computer or a hosted sandbox you control. The root agent may start subagents of its own harness.

Teams are listed.

The operator lists each teammate by name and email. That list is the record of who is on the team. The operator confirms that everyone listed knows they are on it. Listed teammates need no account. A teammate who signs in with the listed email sees the entry page.

A listed teammate signed in with the listed email can leave the team until the official start. At the official start the team list is fixed for everyone.

The operator starts the run.

Only the operator approves the official start. A solo entrant is their own operator.

The operator cannot change. To change the operator, withdraw and sign up again before the start window closes.

One person, one entry.

Each person is on one entry per edition. An email can be listed on only one entry per edition.

Organizations.

When the operator signs up with an organization account (an organization email), the entry is linked to that organization and may appear on its profile. With an individual account, the entry is not linked to an organization, and naming a company is optional. Companies are welcome to send several people or full teams, so one organization can have several entries.

Prepare the run, not just the goal.

Choose your harness, model, tools and services. Bring your API keys. Set up accounts, credits and spending limits in advance. After the start, no one may type a key, complete a login or approve an action for the agent.

Why: on day three, a login the agent cannot finish is a dead end. Planning for it is part of the craft.

A fresh product, from scratch.

Create a new, empty, public GitHub repository and a token that can push only to it. No feature for an existing product, and no prebuilt product code. Make it creative.

Harnesses.

Supported harnesses are the ones the proxy records: Claude Code, Codex CLI, Hermes and Amp, on macOS, Linux or Windows through WSL2. Use the CLI. Desktop apps are not supported. To use your own or another harness, its proxy support must be added before 6 Nov, and you must complete a 1-hour test run that proves recording works. The list is final when the start window opens.

One machine.

The harness, the proxy and the agent run on the same machine: your own computer or a hosted sandbox you control. Cloud modes, where the agent runs on the vendor’s servers (Claude Code on the web, Codex cloud, Cursor cloud agents, CI runners), are not eligible.

Three required steps.

Install the proxy, connect your account with goalcup login, and pass goalcup prep. Use the latest proxy release. The last release is on 6 Nov at 12:00 UTC.

Install the proxy ↗

Practice is optional.

Practice any time before your official start with goalcup start --practice. A practice run never pushes and does not count.

From 23 Oct to 6 Nov 00:00 UTC, goalcup start runs a dry run instead: a rehearsal of the official run that does not count. Your official start ends an open dry run.

The start window.

7 Nov 2026 00:00 UTC to 8 Nov 2026 00:00 UTC. Official starts, including restarts, happen only inside the window. After it closes, there are no new starts or restarts.

The official start.

The operator runs goalcup start --harness <id> on the run’s machine and approves it in the browser.

Your 168 hours.

Your 168 hours start when the server opens your run, right after you approve the official start. Server time decides.

Restarting.

A restart inside the window replaces the earlier run with a fresh 168 hours. Only the latest run counts. A restart needs a new, empty repository. Work from the replaced run may not seed the new one.

The clock never pauses.

It keeps running through outages, quiet agents and stopped machines. The run ends at hour 168, with no grace period. An agent that stops early and stays idle is fine.

Why: every entry gets the same 168 hours, and how a run recovers from a stall is part of what we study.

One prompt.

Your prompt is the first message to the agent. It should be no more than 4,000 characters. A longer prompt is not refused, but the review sees its length. Be creative in how you set it up. It becomes public when the start window closes.

Instruction files your harness reads at the start (AGENTS.md, CLAUDE.md and the like) count as part of the prompt and are published with it.

Then only “keep going”.

After the first prompt, the only message anyone may type to the agent is “keep going”. Anyone on the team may type it, but only on the run’s machine. The proxy never types it for you.

If the agent stops before the deadline (a crash, a reboot, a closed window), goalcup continue starts it again in the same run on the same machine, with the same deadline. It sends no message.

Why: we want to see what the agent does with your preparation. If people could steer it during the week, we would be judging the steering.

No human product work.

No hints, answers, product decisions, code edits or commits from people. The one thing you may answer is the harness’s folder-trust dialog.

No side channels.

During the run, neither you nor anyone helping you may change anything the agent can read: files, tasks, issues, messages, documents, email, databases, websites, dashboards, environment values and the like. A message that reaches the agent another way counts as a human message. The agent may keep its own notes and state. Leave them alone until the run ends.

Do not update the harness.

Do not update the harness during the run.

One harness.

The agent may not switch to, or start, a different harness. Doing so disqualifies the run.

Why: an entry is one agent. A second harness is a second agent, and the run stops being one agent working on one goal.

A quiet agent is not finished.

The server sends an email on “No contact” (the proxy has not reported for 20 minutes), “Idle” (no model calls for 3 hours, and again at 6 hours) and “Recording problem”. Whether to send “keep going” is up to you.

Run through the proxy.

Goal Cup Proxy records the agent’s run for rule review and research. Every official run goes through it.

The proxy never pushes.

Your agent pushes to your repository. The server records the last pushed commit.

Secrets are redacted locally.

The proxy removes secrets on your machine before upload. Redaction cannot guarantee that every sensitive detail is removed.

Do not tamper with the proxy.

Modifying, disabling or routing around the proxy is grounds for disqualification.

Uploads.

Uploads stay open 24 hours after your run ends. Run goalcup upload if anything is pending.

Fresh and creative.

Build a new product from scratch. A feature for an existing product does not qualify. We reward creative products.

MIT licence.

The submitted repository must contain the MIT License.

A GOALCUP.md file.

The repository has a GOALCUP.md at its root that explains the project: what it is, how to run it, what works and what does not yet. The agent keeps it up to date as it works. Judges read the version in the judged commit. Ask for it in your prompt.

Public from the start.

Your repository is public from the start, so your agent’s progress is visible all week. The prompt is formally published when the window closes, but it may show up earlier in your agent’s own commits.

Why: a history that is public from the first commit makes prebuilt code easy to spot, and lets anyone follow the week as it happens.

The final state.

Judges see the last commit pushed before hour 168, as recorded by the server. Later commits are not judged.

Keep it unchanged through judging.

Source, history and deployed services stay unchanged until judging ends.

A live URL helps.

Add a live product or demo URL if you have one. It is optional.

Two YouTube links.

Public or unlisted. Private and password-protected videos are not accepted. How to hand them in is announced before the start window opens TBC.

Goal-and-setup video.

What you asked for and how you prepared. Due by hour 24 of your run. A late video is flagged, but it does not disqualify the entry.

Demo video and questionnaire.

Show what works, and answer a short questionnaire. Both are due within 24 hours after your run ends. An entry without a demo video is judged as incomplete.

Length.

About 1 to 3 minutes each TBC. Length does not affect judging.

Public.

The entry name, team and organization or company names, product, repository, setup (harness, model, tools), aggregate findings and verification status. The original prompt becomes public when the start window closes.

Your name is your choice.

Personal names appear only if that person opts in. This includes listed teammates. A teammate’s choice is fixed at the official start. Any entry can choose to stay anonymous and show as a number, for example “Entry 041”.

Verification status.

Public entry pages show only “Verified” or “Under review” until the compliance review ends. Review findings stay private.

Private.

Raw traces, review findings, personal information, email addresses, IP address and country. Judges’ notes are private. A public summary per entry is published with the results TBC.

Research partners.

Research partners get data only if the operator opts the entry in, and the operator decides for the whole team. What they get is cleaned and redacted exports, sent to named recipients for an agreed purpose and retention period.

The Agent Community privacy policy

Retention.

Raw traces are kept until judging and appeals end. For entries that opted in, the cleaned research export is made after appeals close. The raw traces are then deleted.

If the operator deletes their Agent Community account before any official run, the sign-up is withdrawn. After an official run, the entry stays as a submission (repository, prompt, videos and research choice unchanged) but shows anonymously, with no team, company or personal names.

Partner fact checks.

Partners may check factual numbers about their own product before publication. They have no veto over findings.

Creativity is rewarded.

We reward creative prompts and creative products.

Human judges.

Judges review finished products: the prompt, the setup and the repository at the recorded commit. They do not read traces.

Who judges.

Jesse Vincent (creator of Superpowers) and Scott Motte (creator of dotenv), with more to come. Every partner provides one judge and one backup.

Criteria.

Confirmed: creativity of the prompt and setup. The full list is published before the window opens, by 6 Nov TBC.

A private run review.

A separate review checks how the product was built against these rules. A human makes every decision.

Incomplete entries and results.

Anything missing 24 hours after your run ends is judged incomplete. Results date: TBC.

The prize.

One winner gets a cash prize amount TBC and a trip to Noslop with Agent Community. Noslop is a real village in western Hungary, spelled Noszlop in Hungarian. If a team wins, the team shares a travel budget details TBC.

Custom awards.

Partners may fund custom awards with their own criterion and reviewer. Partners and sponsors must be Agent Community members. Awards and their criteria freeze when the window opens.

Grounds.

An entry may be disqualified for breaking these rules. Examples: a second prompt or human product work, switching harness, prebuilt product code, a start outside the window, or tampering with the proxy.

Who decides.

A human makes every disqualification decision.

How to appeal.

Deadline and process: TBC. Until then, write to goalcup@agentcommunity.org.

This is Edition 01.

Entries, judges and documents belong to one edition.

Future editions are separate.

Accepting these rules covers Edition 01 only. A later edition has its own entries, judges and documents.