01
You reproduce before you conclude
A number that cannot be run twice is not a finding. We would rather ship one defensible result than four suggestive ones.
The work is unglamorous and consequential: benchmark fleets, failure taxonomies, action surfaces, and regulated journeys that have to hold up for both an agent and the person who sent it. Small team, long horizon.
Applications are taken through your own agent, which interviews you and relays what you say. No cover letter. How that works, and[email protected] if you would rather not use one.
Roles are filled as they close, not batched. Nothing here fits: write anyway and say what you would own.
Roughly seventy per cent Python. Published here and, for an agent, at/careers/role.json.
Most of what we do is measured, so most of what we hire for is the habit of measuring. Domain experience in banking or insurance helps and is not required. The instinct to reproduce a result before believing it is.
A number that cannot be run twice is not a finding. We would rather ship one defensible result than four suggestive ones.
Most of our output is read by risk, engineering and marketing at once. Clear prose is part of the engineering, not a wrapper on it.
Regulated change processes, permitted environments and audit trails are the terrain, not friction to complain about.
The last fifth of an agent journey, error recovery, receipts and edge states, is where readiness is actually won or lost.
The job is building surfaces agents have to use, so we would rather watch you work with one than read a cover letter about it. The agent asks, you answer, it relays. About forty minutes.
One command connects your agent to Webzero, with nothing to install and nothing to download. It walks you through eight questions, reading each one to you and recording what you say.
Your agent is a scribe, not a ghostwriter. It may explain a question, let you think out loud and fix your spelling. It may not write your answer, and it is told so at every step.
Every answer carries a declaration of who authored it, is timed against when the question was issued, and is read for whether a person or a model produced it. An application the agent answered is marked and does not go into the pile.
Nobody is disadvantaged for not using an agent. Write to [email protected] with the role in the subject line and it reaches the same person.
claude mcp add --transport http webzero-apply https://www.webzero.ai/careers/mcpAny MCP client works: the transport is http and there is nothing to install. The skill your agent reads first is at/skills/apply-to-webzero/SKILL.md. Then tell it you want to apply.
Eight questions, answered by you, relayed by your agent. Or one email with the role in the subject line. A person reads every application either way.