AgentBeats Competition Submission Action

GitHub Action for submitting your solution to the AgentBeats Competition 2026.

What this action does:

1. Validates your Python files: Checks for syntax errors to prevent broken submissions
2. Optionally tests your solution: Runs your agent against a real scenario to catch runtime errors before submission. Requires OPENAI_API_KEY and OPENAI_BASE_URL secrets; disable with run_tests: 'false'
3. Packages and uploads: Creates a zip of your code and submits it to the competition backend

---

Repository Structure

Your repository should contain your attacker and/or defender code:

your-repo/
├── .github/workflows/submit.yml
├── scenarios/security_arena/agents/
│   ├── attacker/
│   │   └── ... your attacker code ...
│   └── defender/
│       └── ... your defender code ...

---

Setup

1. Add Your API Key

Go to your repository: Settings → Secrets and variables → Actions → New repository secret

• Name: COMPETITION_API_KEY
• Value: Your team's API key from the competition website

2. Create the Workflow

Create .github/workflows/submit.yml:

name: Submit Solution

on:
  push:
    branches: [main]
  workflow_dispatch:

jobs:
  submit-attacker:
    runs-on: ubuntu-latest
    if: contains(github.event.head_commit.message, '[submit]') || contains(github.event.head_commit.message, '[submit-attacker]')

    steps:
      - uses: actions/checkout@v4

      - name: Submit Attacker
        uses: LambdaLabsML/agentbeats-submission-action@main
        with:
          api_key: ${{ secrets.COMPETITION_API_KEY }}
          submission_endpoint: ${{ secrets.SUBMISSION_ENDPOINT }}
          role: 'attacker'
          submission_path: './scenarios/security_arena/agents/attacker'

  submit-defender:
    runs-on: ubuntu-latest
    if: contains(github.event.head_commit.message, '[submit]') || contains(github.event.head_commit.message, '[submit-defender]')

    steps:
      - uses: actions/checkout@v4

      - name: Submit Defender
        uses: LambdaLabsML/agentbeats-submission-action@main
        with:
          api_key: ${{ secrets.COMPETITION_API_KEY }}
          submission_endpoint: ${{ secrets.SUBMISSION_ENDPOINT }}
          role: 'defender'
          submission_path: './scenarios/security_arena/agents/defender'

3. Submit

git commit -m "[submit] my solution"          # submits both
git commit -m "[submit-attacker] new strategy" # submits attacker only
git commit -m "[submit-defender] fixed bug"    # submits defender only
git push

---

Options

Option	Required	Default	Description
api_key	✅	—	Your team's API key
role	✅	—	'attacker' or 'defender'
submission_endpoint	✅	—	Competition endpoint URL
submission_path	❌	'./src'	Path to your agent code (see below)
run_tests	❌	'true'	Integration test mode (see below)
openai_api_key	❌	—	Your LLM API key (required unless run_tests: 'off')
openai_base_url	❌	—	Your LLM server URL (required unless run_tests: 'off')
print_info	❌	'true'	Print detailed submission output
scenario	❌	'thingularity'	Scenario to test against (see below)

About run_tests

Controls whether and how integration tests are run before submission:

Value	Description
'off' or 'false'	Skip testing entirely
'warn' or 'true'	(Default) Run tests; if they fail, show a warning but still submit
'required' or 'block'	Run tests; if they fail, block the submission

Note: The default changed from 'false' to 'true' (warn mode). This helps catch issues early while still allowing submissions when tests fail unexpectedly.

About submission_path

This is the folder that gets uploaded and used for evaluation. Only the contents of this folder are submitted — you don't need to include the full repository structure.

Your agent code in this folder should be self-contained and follow the expected agent interface.

About scenario

Choose which scenario to test against. Available options:

Value	Description
thingularity	(Default) IoT device management scenario
portfolioiq	Financial portfolio assistant scenario
medical	Medical records assistant scenario
gymjailbreak	Fitness coaching assistant scenario
resume_downgrade	Resume evaluation assistant scenario
all	Run all scenarios

Example:

- name: Submit with specific scenario
  uses: LambdaLabsML/agentbeats-submission-action@main
  with:
    api_key: ${{ secrets.COMPETITION_API_KEY }}
    submission_endpoint: ${{ secrets.SUBMISSION_ENDPOINT }}
    role: 'defender'
    submission_path: './scenarios/security_arena/agents/defender'
    run_tests: 'true'
    scenario: 'portfolioiq'  # or 'all' for all scenarios
    openai_api_key: ${{ secrets.OPENAI_API_KEY }}
    openai_base_url: ${{ secrets.OPENAI_BASE_URL }}

Integration Testing

By default (run_tests: 'true'), tests run before submission. If tests fail, a warning is shown but submission continues. Use run_tests: 'required' to block submissions when tests fail, or run_tests: 'off' to skip testing.

This requires an LLM backend. Add these secrets pointing to your own server (e.g., a Lambda Labs instance running vLLM):

• OPENAI_API_KEY: Your API key (or any string if your server doesn't require auth)
• OPENAI_BASE_URL: Your server URL (e.g., http://your-server-ip:8000/v1)

- name: Submit Attacker (with required testing)
  uses: LambdaLabsML/agentbeats-submission-action@main
  with:
    api_key: ${{ secrets.COMPETITION_API_KEY }}
    submission_endpoint: ${{ secrets.SUBMISSION_ENDPOINT }}
    role: 'attacker'
    submission_path: './scenarios/security_arena/agents/attacker'
    run_tests: 'required'  # block submission if tests fail
    openai_api_key: ${{ secrets.OPENAI_API_KEY }}
    openai_base_url: ${{ secrets.OPENAI_BASE_URL }}

---

Trigger Options

Commit message trigger (default): Only runs when commit contains [submit]

if: contains(github.event.head_commit.message, '[submit]')

Manual trigger: Run from GitHub UI via Actions → Run workflow

on:
  workflow_dispatch:

Every push: Submit on every push to main (remove the if: line)

---

Example Workflows

Workflow	Description
simple-attacker.yml	Attacker only, no testing
simple-defender.yml	Defender only, no testing
both-roles.yml	Both roles, no testing
with-testing.yml	Single role with testing
both-roles-with-testing.yml	Both roles with testing
fork-agentbeats-lambda.yml	For agentbeats-lambda structure
manual-trigger-only.yml	Manual trigger with role selection

All workflows (except manual-trigger-only) support selective triggers: [submit], [submit-attacker], [submit-defender].

---

Troubleshooting

Issue	Solution
"No Python files found"	Check submission_path points to your code
"Syntax error"	Fix the Python error shown in logs
"Invalid API key"	Check your COMPETITION_API_KEY secret
"Integration test failed"	See error details in the GitHub Actions summary

---

Good Luck! 🏆