Harness Engineer

We turn local creators into the marketing engine for the businesses in their own neighborhood. Creators who live nearby come in, eat, film, and post.

Owners tell us people walk in every day saying "we saw you on TikTok." Almost half our new restaurants come from other owners referring us.

AI agents run the operation end to end. They match creators to restaurants, book the visits, track the content, and handle all conversations over iMessage. Hundreds of visits a month, almost none of it touched by a human.

We're live with restaurants in Southern California and growing 40% month over month.

Most of our users never really touch our app. They text us. Restaurant owners and creators handle almost everything through conversation, and on our side that's a set of AI agents doing the booking, the follow-ups, the scheduling, and the problem solving.

You own those agents. Not the prompts alone, the whole system underneath them: how they understand what someone actually wants, how they call tools without breaking, how they hold context across a conversation that's been going for 6 months, and how they take real actions in the real world without us having to check their work.

Any text that comes in, your systems handle it.

The agents. Full ownership of both of our agents end to end. How they're built, what they can do, and what happens when they get something wrong.

The harness. The layer coordinating LLMs, tools, memory, and async workflows. This is the actual engineering problem here and it's most of your week.

Tooling that doesn't fall over. Schema validation, retries, permissions, error handling. An agent that calls a tool correctly 95% of the time is not good enough when it's booking real visits at real restaurants.

Evals. Frameworks that measure task completion, accuracy, safety, latency, and failure modes. If we change a prompt on Tuesday, we should know by Tuesday whether it made things worse.

Observability. Traces, logs, and failure analysis across agent workflows. When something goes wrong in a conversation, you should be able to see exactly where.

Turning vague into dependable. A restaurant owner texts something ambiguous at 11pm. Getting from that to a reliable agent behavior is the hard part, and it's the part we care most about.

Required

• 2+ years building software, with real experience building agent systems or harnesses. Not just calling an API in a side project
• Strong conversational agent experience. This is the thing we weigh most heavily. Our product is a conversation, and someone who has only built single-turn or task-runner agents will struggle here
• Strong database and system design
• You've shipped agents that real people used and dealt with the fallout when they broke
• Comfortable with ambiguity. There's no established playbook for most of this
• Based in Orange County / willing to relocate and able to work onsite
• This isn't a 9-to-5

Nice to have

• iMessage agent experience
• You've built eval frameworks, not just run them
• Observability and tracing work on LLM systems
• Restaurant industry or creator economy experience

We're building long-horizon conversational agents that operate in the physical world.

You'll work across Postgres, REST APIs, multi-agent harnesses, async workflows, memory management, and iMessage and SMS infrastructure. Conversation is our main product surface, so messaging is core infrastructure here rather than a channel we bolted on.

We work heavily with AI coding tools and expect you to as well. Your work will span backend infrastructure, agent systems, and the apps our restaurants and creators use. You'll own architectural decisions and take ideas from prototype to production.