{
  "$comment": "The open engineering role at Webzero, as one document. Static, no credential, Access-Control-Allow-Origin: *. Reading this is the whole journey: there is nothing to call and nothing to register for. Applying is a separate thing and needs the candidate, see apply.",
  "updated": "2026-09-12",
  "publisher": "Webzero",
  "publisherUrl": "https://webzero.ai/",
  "roles": [
    {
      "id": "product-engineer",
      "title": "Product Engineer",
      "team": "Engineering",
      "employment": "Full-time",
      "location": "Remote-first",
      "url": "https://webzero.ai/careers",
      "summary": "Build and run the surfaces Webzero measures agents against: the scoring engine behind WARI, the sandboxed agent fleet behind Agentrek, the consoles both are read through, and the agent-facing files on webzero.ai itself. Python-first, roughly seventy per cent backend.",
      "stack": [
        "Python 3.11, FastAPI, async SQLAlchemy and SQLModel, Alembic, Postgres in production and SQLite locally",
        "Playwright, the Anthropic SDK, and MCP: a client in WARI, a stdio server in services/, WebMCP in the browser",
        "Docker and ECS Fargate, Redis, S3",
        "The WARI and Agentrek consoles are Jinja templates, hand written vanilla JavaScript and CSS custom properties. There is no frontend framework in either console.",
        "The marketing site is Astro with React 19 islands and TypeScript",
        "Small Node services with zero dependencies"
      ],
      "looking_for": [
        "Async Python at a senior level, including the ways a shared session or an unclosed pool actually fails in production",
        "Has built or broken an LLM agent: tool use, structured outputs, MCP, a real browser under Playwright",
        "Comfortable without a frontend framework, and comfortable dropping into Astro and React when the marketing site needs it",
        "Container and cloud fluency as a user: can debug a Fargate task that died, not asked to design a platform",
        "Writes. Most of what we ship is read by risk, engineering and marketing at once, and the written rules are the ones that hold",
        "Reads prompt injection and agent-facing surfaces as a security problem, because here they are one"
      ],
      "not_required": [
        "React ecosystem depth, Next.js, GraphQL, Kubernetes, model training. None of it is in the codebase."
      ]
    }
  ],
  "apply": {
    "how": "Through your own agent. Webzero publishes an Agent Skill and an MCP endpoint your agent connects to; the agent interviews you, relays what you say, and submits it. It is not a form an agent fills in on your behalf: the answers have to be yours.",
    "skill": "https://webzero.ai/skills/apply-to-webzero/SKILL.md",
    "requires": "A person. The application is verified by an emailed code and the answers are checked for who wrote them.",
    "without_an_agent": "Write to team@webzero.ai with the role in the subject line. Nobody is disadvantaged for not using an agent.",
    "note": "The endpoint that takes applications is hosted by Webzero and speaks MCP over POST; nothing runs on the candidate's own machine. It is a hiring intake and not a product surface: it measures nothing, prices nothing and starts no run. The w0 score console and Agentrek are not publicly callable, and reading this file needs no token, account or registration."
  }
}
