About the Opportunity

Supabase — the Postgres development platform used by 7M+ developers — is building an AI-native internal operating system: a common way of working where AI carries a meaningful share of the operational load, not just sits beside it as an assistant. A new central team is standing this up, and you'd be the engineer on that team.

The Mission

You own the execution layer: the platform that actually runs agents. An event-triggered queue, a headless model-agnostic runtime (an open architecture decision you'll close), durable state that survives a restart, a human review gate, atomic rollback, and complete logging of every prompt, tool call, and decision so any run can be reconstructed.

What You'll Build

The evaluation layer — golden suites with behavioral assertions, written judge rubrics, safety cases that pass on every run, and a CI gate that blocks regressions. An agent that can't be measured can't be trusted.
The agent portfolio — reporting, drafting, linting, triage, and Q&A agents across executive, team-lead, and IC layers, plus a meta layer that watches the platform and improves it.
Governance in code — risk tiers the pipeline actually enforces, least-privilege credentials per agent, tool-permission gates, a decision audit log, and autonomy classes where the dangerous class has no code path at all. An agent may read and report, carry a human-approved update into a system of record, and initiate contact within a strict budget — but it must be structurally impossible for it to autonomously write a commitment into a shared work system.
The human-touch design — a hard interruption budget per person, message bundling instead of a ping stream, and a structure that gives something useful before it asks.
The platform's own ROI ledger — a pipeline that logs the work each agent absorbs and computes a monthly figure, so value is a measurement, not a claim.

How You'll Think

Recursive: build the generator, not the artifact. When you need thirty agents, you build the inventory, compiler, and distribution path that makes the thirty-first cost an afternoon. Inversion: start from the failure and work backward — the credential in the tool grant physically can't set an owner, due date, or status, so no prompt injection or clever input can ever force it.

Why Join

Fully remote with global hiring, ESOP for every team member, tech allowance, and 100% health coverage (80% for dependents). This is a rare greenfield mandate: you close architecture decisions yourself and own the infrastructure end to end.

🚀 Ready to Apply?

👉 Click Here to Apply Directly for this Position

Similar Roles

Data & AI

PositionLocation

Staff Data Scientist, Consumer Growth — Reddit (Remote, Ontario Canada)

Reddit•Remote, Ontario Canada

Data & AI6d ago

Senior Data Scientist, Demand Forecasting — Flix, Munich

Flix•Munich, Germany

Data & AI6d ago

Senior Marketing BI Analyst (Referral & Virality Analytics) — Berlin, Germany

HelloFresh•Berlin, Germany

Data & AI6d ago

Bank Relationship & Analytics Manager (Fintech, Savings) — Berlin, Germany

Raisin•Berlin, Germany

Data & AI6d ago

About the Opportunity

The Mission

What You'll Build

The evaluation layer — golden suites with behavioral assertions, written judge rubrics, safety cases that pass on every run, and a CI gate that blocks regressions. An agent that can't be measured can't be trusted.
The agent portfolio — reporting, drafting, linting, triage, and Q&A agents across executive, team-lead, and IC layers, plus a meta layer that watches the platform and improves it.
Governance in code — risk tiers the pipeline actually enforces, least-privilege credentials per agent, tool-permission gates, a decision audit log, and autonomy classes where the dangerous class has no code path at all. An agent may read and report, carry a human-approved update into a system of record, and initiate contact within a strict budget — but it must be structurally impossible for it to autonomously write a commitment into a shared work system.
The human-touch design — a hard interruption budget per person, message bundling instead of a ping stream, and a structure that gives something useful before it asks.
The platform's own ROI ledger — a pipeline that logs the work each agent absorbs and computes a monthly figure, so value is a measurement, not a claim.