Agent Harness Engineer

Top-of-the-market salary and the kind of ownership that only exists at this stage.

This role is remote-first, with hubs in Munich, New York, and Warsaw. We bring the whole team together in person a few times a year.

• You've personally built or significantly modified an AI agent harness, and can describe the trade-offs you made.

• You understand how an agent's system prompt, skills, and tools interact on diverse and complex problems, and you've tuned that interplay yourself, not just read about it.

• You've built custom skills, commands, CLIs, or MCP servers to make your own agentic coding faster.

• Agentic coding is your daily workflow, and you've stress-tested models and harnesses across the frontier.

• Systems thinking. You make hard technical trade-offs and design for how things break at scale.

• Speed, with the bar up. You ship today, not Thursday.

• Genuine interest in how AI works. You've read the papers and have opinions about evals.

• Remote. Work from wherever you're sharpest; we default to async and gather in person a few times a year.

Python. No agent framework, no orchestration library. We build the harness ourselves, from the loop up.

• Previous AI engineering experience: agents, harnesses, eval infrastructure.

• Previous founder or early-stage builder.

• Open source contributions to the AI tooling we live in.

You're the person who makes Viktor do more things, for more customers, more reliably. You've built agents before (runtime, tools, memory, evals) and have opinions about what makes them work. You ship the day you write the code and reach for agentic engineering by default. If you've never built an agent, this isn't the role.

We're building a harness ready for AGI. Models keep getting better on their own; the harness decides how much of that intelligence turns into real work. Viktor shouldn't wait to be handed tasks. It should explore what a company is doing and automate it on its own.

Getting there is context engineering at the frontier: how the system prompt, skills, memory, and tools interact on diverse, complex problems, and how that interplay holds up across the real work of 56,000+ workspaces.

• No layers. You work directly with both founders, and decisions get made in the room, not in a Linear ticket.

• The agent is the product. Every reliability win shows up in retention the next week.

• Frontier work. You're building a harness at the edge of what's currently possible, experimenting with agents that can take on full responsibilities, not just tasks.

• Build the agent runtime: the loop where it plans, acts, and verifies its own work instead of declaring success it didn't earn.

• Engineer the context: shape how the system prompt, skills, memory, and tools combine, and how the agent holds up when problems get diverse and complex.

• Make the agent fluent with tools: composition across many integrations, recall, and the skill files that keep the right functions at hand.

• Grow what the agent can do: memory it can trust, skills it writes for itself, work it runs on its own.

• Make it measurably better: evals, catching regressions, and turning every failure into a fix.

• Ship product: features that reach users through Slack and Teams within hours.

• Whatever needs building. Small team, large surface.