Designerhuman–agent interfaces
The problems you’ll own
- How do you show a person a thousand agent actions in two minutes without hiding the one that matters?
- How do you make an AI’s decision readable at a glance: its evidence, the precedent it followed, what it rejected, how sure it is?
- What does “act, ask, never” look like, so a busy executive sets it correctly the first time and trusts it afterwards?
- When should the system interrupt a person, and when should it stay quiet?
What you’ll do
- Design the daily brief, the decision record, the authority controls and the token ledger, from first sketch to shipped product.
- Prototype at working fidelity and test with the operators who run agent teams.
- Own the company’s visual language, from the product to the figures in our research.
You might be a fit if
- You have shipped dense, data-heavy products such as developer tools, analytics or financial software.
- Information design is your strength, and you can defend a design decision with evidence.
- You would rather build the prototype yourself than describe it.
AI Architectthe management layer
The problems you’ll own
- How do twelve agents share one picture of the world without drowning each other in context?
- What consistency model fits a shared state that many agents write to at once, and what happens when two of them commit plans that cannot both be true?
- How do you store a decision so it can be replayed faithfully years later, against the definitions that were believed at the time?
- How do you grant authority in tiers and keep every automated action reversible?
- How do you tie every token to an outcome, across agents we did not build?
What you’ll do
- Own the architecture of the management layer: shared state, context routing, the decision record, authority and rollback, cost accounting.
- Define the interfaces that let any vendor’s agent be managed, reviewed and taught.
- Build the replay and non-regression harness that every learned change must pass.
- Work with our researchers to take methods from a paper to production.
You might be a fit if
- You have designed and operated distributed systems in production, and can talk about how they failed.
- You have built agent systems that went past the demo: multi-agent, long-running, tool-using.
- You are strong on data modelling and consistency, and you care about observability.
- You write clearly enough that others can build on your decisions.
Researcherorganizational intelligence
The problems you’ll own
- How do you estimate the value of a decision policy when outcomes arrive 60 days late and are confounded by everything else the company did?
- How do you assign credit across agents that produced a result together?
- How does a system know when not to decide, and stay calibrated as the business drifts?
- How does one agent’s lesson transfer to the others without breaking what they already do well?
- When is an expert’s minute worth more than another thousand tokens?
What you’ll do
- Take one of the six problems in the Super Intelligence Lab and own it end to end: formalize it, build it, measure it on real organizational data.
- Design the measurement that could prove us wrong, then run it.
- Write up what holds, including the negative results.
You might be a fit if
- You have a research record in one of: causal inference and off-policy methods, multi-agent learning, continual learning, uncertainty quantification, human–AI decision making, or knowledge representation.
- You implement your own ideas.
- You prefer messy real data to a clean benchmark.