Member of Technical Staff - Product Engineering (US)

Being a Member of Technical Staff at Basis means you'll face changing problems. You might spend a week designing the context ontology for an agent, then shift to building the data pipelines to hydrate agent state.

We don't have static functional teams. We have pods that exist around areas and objectives which reform every quarter as our goals update and the team expands. It's common to see engineers do core infra work one quarter and product agent work the next. It's also common for engineers to deeply specialize in certain technical areas.

As a Member of Technical Staff, you can do work across:

Product Engineering

• Agent Engineering (e.g. context engineering, tool design)

• Agent Platform (e.g. harness engineering, eval systems)

Platform & Infrastructure

• Agent Data (e.g how we ensure accurate & low latency external data for agents)

• Atlas, the mandate to build the AI systems powering Basis from recruiting to sales to engineering. The Atlas team builds the context layer, internal agents, and knowledge systems that will eventually produce the majority of total output at Basis.

Some problems we think about

How do we ensure users feel as comfortable working with nondeterministic systems (agents) as they do with deterministic software?

• We say around here that "users are already used to working with non-deterministic systems, it's just those systems are their coworkers, not their computers."

How do we manage agent state over long horizons?

• What should ETL workflows look like when agents operate for hours? How should state hydration work? How much agency do we grant the agent to decide how its own data state is managed?

At what level of abstraction do we design tools for agents?

• Too low and the agent wastes effort on mechanical steps. Too high and it can't handle edge cases. The right answer changes as models get smarter.

How do we extract dense signal from long-running agent trajectories?

• An agent runs for five hours across thousands of decisions. How do we attribute outcomes back to specific reasoning steps? How do we tune eval judges when the judgement includes subjectivity?