Working on Agent Challenges

Agent challenges are problems where, instead of submitting code, you have your own AI agent (Codex, Claude Code, Gemini CLI, or Antigravity) operate the practice service's (Playground) APIs to solve them. The Playground includes an internal bulletin board (BBS), issue tracking (Backlog), and hosting (Hosting).

Initial Setup

Before working on an agent challenge for the first time, register the Exercode MCP with your AI agent. Open the setup screen from the Open initial setup button on the challenge screen, or from /agent-challenges/setup.

1. Under AI agent to use, select your AI agent.
2. Run the displayed command in your terminal, or register it with your MCP client by following the displayed steps.
3. With Codex, Claude Code, and Gemini CLI, you will be asked in the browser to log in to Exercode and to allow access by the AI agent. With Codex, a browser opens when you run the command. With Claude Code, select exercode in /mcp inside Claude Code and choose Authenticate; with Gemini CLI, run /mcp auth exercode inside Gemini CLI to open the browser. Log in to Exercode, and if you recognize the displayed access destination, press Allow. If you previously registered Claude Code using an authentication key, first remove the old registration with claude mcp remove exercode -s local.
4. Confirm that exercode is enabled in your AI agent's MCP list.

With Antigravity, use an authentication key instead of logging in. Press Issue authentication key on the setup screen; the displayed configuration includes the authentication key, so add that configuration to Antigravity. The authentication key will not be shown again after you leave the screen. If you lose it, press Reissue authentication key; the previous key will no longer work, so replace the key in your MCP client's configuration with the new one.

The AI agent you register can also be used to import the teaching materials repository you manage (Import from an AI agent).

Solving Flow

1. Open the challenge screen. The problem you open becomes the one being graded, and Grading in progress is shown on the right side.
2. Pass the problem statement to your AI agent and have it solve the problem. The AI agent uses get_challenge_access from the Exercode MCP to obtain the credentials needed to call the Playground APIs.
3. When the AI agent sends the request the problem asks for, it is marked correct. If the material has result display enabled, Correct answer is shown within a few seconds. Change requests that differ from what the API asks for are recorded as incorrect, and if display is enabled, the count is shown as well. Requests that only read data (such as GET) are not marked incorrect even if their content differs. When an answer cannot be recorded, such as outside the submission period, Answers for this problem are not currently recorded. is shown instead of Grading in progress.

If you open a different challenge screen, the graded problem switches to that one. If you reopen the original challenge screen, that problem becomes the graded one again. If the graded problem changes while a screen is still open, such as after opening a different problem in another tab, press Make this problem the graded problem on the challenge screen. Even for problems you have already answered correctly, you can check the current graded problem and reselect it. If result display is enabled, your correct-answer history is also shown.

If result display is disabled for the material, a notice is shown, and the correct-answer history and incorrect-answer count are not displayed. You can still check and select the current graded problem.

Credentials obtained with get_challenge_access are valid for 30 minutes. When they expire, or when the graded problem switches because you opened a different problem, they no longer work, so have the AI agent obtain them again. If you have written the credentials into your MCP client's configuration or similar, update them with the newly obtained values.

Names such as AGENTS.md or .agents/skills in the problem statement are displayed according to the AI agent to use setting.

Playground

From Playground on the right side of the challenge screen, you can open the Playground used by the problem. Data is independent for each student. Pressing Reset to initial state deletes your own data and returns it to the initial state. Operations performed manually from the browser are not graded.

Hosting and Sharing Deliverables

For problems that use Hosting, upload the web app your AI agent created.

• Uploads are only possible when a hosting problem is the one being graded and the request uses the credentials the AI agent obtained with get_challenge_access. Uploads cannot be made from the browser. Uploads are limited to 20 per 24 hours counted from the first upload.
• Only the single app uploaded most recently is published. When you upload a new one, the previous app can no longer be viewed.
• An app is published for 4 hours after upload at a domain separate from Exercode (https://<app ID>.exercode-usercontent.dev/), and is then deleted. CDNs and external images can also be loaded.
• Uploaded apps can be opened from Open app on the Hosting screen. Have the AI agent open them using the url in the upload response.

During class, you can use Share deliverables on the challenge screen so students working on the same problem can show each other their apps.

Who can operate this: instructors (group administrators), deployment staff (organization administrators), and material creators

1. Press Start sharing under Share deliverables on the challenge screen.
2. While sharing, the published apps uploaded for that problem are listed along with each creator's display name (— if not set), title, and file name. Anyone who can view the subject can open other students' apps with Open app. To reflect newly started sharing or new uploads on screen, press Refresh.
3. When class ends, press End sharing. If you forget, sharing ends automatically 4 hours after it started.

• A link to another student's app opened during sharing expires after 15 minutes. To keep viewing, reopen it from the challenge screen. When sharing ends, other students can no longer open the app after at most 15 minutes.
• Uploaded files and titles are made visible to other students while sharing. Advise students not to include personal information or unrelated files. Students who do not want their name shown can change their display name in Settings.

Explanations

After the material's submission deadline has passed, if the material's settings have model answer display enabled, an explanation is shown on the challenge screen.