⚠️ Please read first

• This is a full-time, in-person role based in San Francisco (Presidio). We work from the office 5 days a week.

• You must be based in the Bay Area or willing to relocate before starting.

• We require US work authorisation, but are open to O-1 or J-1 visa sponsorship for exceptional candidates.

About Sauna by Wordware

Every tool you use was built as an island, so you became the glue between them: a name from Gmail into the CRM, a time from the CRM into Calendly, a note from Calendly into a doc. AI showed up and mostly became an eleventh window to paste into.

Sauna is a hybrid of AI and software standing on one shared brain. It holds your context and hands you real software built on top of it, so an ordinary Calendly becomes one that writes the invite itself and knows which of your busy hours bend, and for whom. Software for the path you repeat, AI for the edge that never repeats.

Wordware is the team behind it, 15 people backed by Spark Capital, Felicis, and Y Combinator with a $30M seed, the largest in YC's history. We build from an office in the Presidio, fifty metres from Crissy Field beach, and we work absurdly hard because we can see and feel the outcome every day.

About the Role

Sauna's agent runtime has to keep track of work when a tool fails or new code ships mid-task. One session might start with a message, check Calendar, draft a Slack reply, and update Linear after the first response. A tool call can time out after the action succeeds; the runtime needs to know what happened before retrying.

You'll build the execution engine behind those sessions, choosing when work belongs in an actor and when a job is enough. You'll also design how long-running processes deploy while other tasks are in flight. This is a senior, hands-on engineering role. You'll make those calls with the team, write the code, and own it in production. We'll teach you our agent harness.

What You'll Do

• Build and improve the runtime that starts agent work, carries session state through tool calls, and resumes interrupted runs.

• Choose between actors and jobs for different kinds of work. Implement the choice and check how it behaves when sessions overlap.

• Design how long-running processes drain or resume when code changes, keeping in-flight work accounted for across releases.

• Handle retries and timeouts around tool calls, including cases where an external action succeeded but the response never came back.

• Make runtime failures traceable. Find why a run stopped or stalled, fix the execution path, and see the change through to production.