Founding Engineer — Agent Harness
Il y a 8 heures
Paris, Ile-de-France
Rowads Studio
Temps plein
Gratuit avec email ou Google
Enregistrez cette offre et organisez votre recherche
Créez un compte gratuit pour enregistrer des offres d'emploi, créer des alertes et revenir à cette liste depuis votre tableau de bord.
Gratuit avec email ou Google
En continuant, vous acceptez nos Conditions d’utilisation & Politique de confidentialité.
Founding Engineer — Agent Harness (Extra)
📍 Paris, on-site
· CDI or freelance
· Reports to the CTO 7:02 am. An agent wakes up from a timer it set three days ago. It picks up exactly where it left off, sees the contract is now signed, files it, and goes back to sleep. Nobody pressed a button. Two tasks over, another agent is fixing a bug in the platform it runs on, and shipping the fix. That's Extra. We're looking for the person who makes it bulletproof. The models are good enough now. Whether an agent finishes the job, or burns €40 and stalls at step two, depends on everything around the model: the harness. That's the job. 🧰 What Extra is Every task is a Claude agent in its own cloud container. Around it, Extra gives the agent:
• shared drives;
• memory that carries from one task to the next;
• a real browser;
• a vault that lets it use your Gmail or GitHub without a secret ever touching the container. Rowads, a Paris creative studio, already runs on Extra for HR paperwork, contracts, video reviews and ad reporting. Now we're turning it into a product. You'd be its first dedicated engineer, working directly with the CTO. Stack: TypeScript on Cloudflare (Workers, Durable Objects, Containers), Claude, and MCP everywhere. 🔥 Problems you'll only find here
• Agents nap. Containers hibernate after five idle minutes, and background jobs, sub-agents and timers all have to survive it. "Done" has to mean done.
• Context is the bill. On one real task, 71% of the cost was the agent re-reading its own conversation, and 5.6% was the model actually writing. Fix that ratio.
• Memory should help, not haunt. What to remember, for whom, and when to forget it.
• Model routing nobody has built yet. The right model for each step, while every new release quietly breaks one of your assumptions.
• The attacker is often the content. A web page, a PDF, or a message from another agent can carry the attack, while the agent holds real credentials. Prompt injection and blast radius are everyday design problems here. 🎯 What you'll own You'll own reliability, security, context, memory and model routing, end to end. There's no ticket queue: you watch where agents trip users up and decide what to build. Within a year, you'll be one of a handful of people who really understand agent harnesses. ✅ You'll love it here if
• You've run production systems at senior level, and been the one paged when they broke.
• You've built something real with LLM agents, beyond calling a chat API.
• You think in failure modes, blast radius and cost per task.
• You read model release notes the day they come out. 🚫 You won't if
• You want a big team, a career ladder, or a spec before you start.
• You'd rather not write code with agents. We build Extra with Extra. ✉️ Apply with something you've built with agents, plus the strangest thing you've ever seen one do. A CV is optional.
· CDI or freelance
· Reports to the CTO 7:02 am. An agent wakes up from a timer it set three days ago. It picks up exactly where it left off, sees the contract is now signed, files it, and goes back to sleep. Nobody pressed a button. Two tasks over, another agent is fixing a bug in the platform it runs on, and shipping the fix. That's Extra. We're looking for the person who makes it bulletproof. The models are good enough now. Whether an agent finishes the job, or burns €40 and stalls at step two, depends on everything around the model: the harness. That's the job. 🧰 What Extra is Every task is a Claude agent in its own cloud container. Around it, Extra gives the agent:
• shared drives;
• memory that carries from one task to the next;
• a real browser;
• a vault that lets it use your Gmail or GitHub without a secret ever touching the container. Rowads, a Paris creative studio, already runs on Extra for HR paperwork, contracts, video reviews and ad reporting. Now we're turning it into a product. You'd be its first dedicated engineer, working directly with the CTO. Stack: TypeScript on Cloudflare (Workers, Durable Objects, Containers), Claude, and MCP everywhere. 🔥 Problems you'll only find here
• Agents nap. Containers hibernate after five idle minutes, and background jobs, sub-agents and timers all have to survive it. "Done" has to mean done.
• Context is the bill. On one real task, 71% of the cost was the agent re-reading its own conversation, and 5.6% was the model actually writing. Fix that ratio.
• Memory should help, not haunt. What to remember, for whom, and when to forget it.
• Model routing nobody has built yet. The right model for each step, while every new release quietly breaks one of your assumptions.
• The attacker is often the content. A web page, a PDF, or a message from another agent can carry the attack, while the agent holds real credentials. Prompt injection and blast radius are everyday design problems here. 🎯 What you'll own You'll own reliability, security, context, memory and model routing, end to end. There's no ticket queue: you watch where agents trip users up and decide what to build. Within a year, you'll be one of a handful of people who really understand agent harnesses. ✅ You'll love it here if
• You've run production systems at senior level, and been the one paged when they broke.
• You've built something real with LLM agents, beyond calling a chat API.
• You think in failure modes, blast radius and cost per task.
• You read model release notes the day they come out. 🚫 You won't if
• You want a big team, a career ladder, or a spec before you start.
• You'd rather not write code with agents. We build Extra with Extra. ✉️ Apply with something you've built with agents, plus the strangest thing you've ever seen one do. A CV is optional.