SWE - Agent Engineer

About Blue: Blue is building a new kind of phone, with your chief of staff built in. Blue sees what you do and reads every notification, text, and email for you. Based on everything it’s learned about you, it decides what to do with each one: handle it and leave you an FYI, or draft a reply and wait for your OK. Anything in your life is one ask away from being off your plate. Blue can manage complex long term tasks across apps

It works from day one. When you switch, Blue reads everything that comes over from your iPhone, so it already knows the people in your life, how you write, and the passwords, cards, and IDs it needs, kept in a secure vault you both share. No onboarding, and no explaining yourself.

Delegation this deep isn’t possible from a single agent in an app. It takes owning the operating system, the whole intelligence stack, and, ultimately, the hardware. Today the Blue OS runs on Google Pixel hardware while our own hardware is in development. We’re Y Combinator-backed, seed-funded, going into our series A, and our first customers already use it as their daily phone.

The Role:
Blue's assistant works in two modes: live, when you talk to it, and in the background, when it handles your notifications, texts and email and takes work off your plate across apps and services. You will make it more capable and more reliable, and prove both with measurements rather than demos.
This role is open at mid and senior levels, including a lead who will own the background assistant's architecture and lead a small team. Depending on your background, you will own some of the areas below.

What you'll own

• Background assistant architecture: How tasks are planned, split across sub-agents, handed off, run and recovered
• Evals, quality and cost: The evals that tell us whether a change made Blue better or worse, quality tracking from real use, LLM spend per user, and the consented, redacted record of agent runs that our own models will train on
• Connectors and capabilities: Connecting Blue to APIs, MCP servers and private endpoints, and new abilities such as image generation, hosting web pages, writing code and sending PRs, and making phone calls
• Screen control and memory: The tools Blue uses to operate apps on screen, races between a changing screen and the model's decision, scripts or small models for tasks it has done before, and how it remembers and recalls what it knows about you
• Safety and review: Confirmation before risky actions, permissions, and resistance to prompt injection; reviewing agent changes from teammates and AI agents
• Cross functional planning: You will work with the ML engineering side and the product side to plan the implementation and delivery of assistant capabilities.

What we're looking for

• You have shipped LLM agents or tool-using LLM features to real users
• You measure: you have built evals and trust them more than a good demo
• You understand the security risks of agents acting for a user, especially prompt injection
• You do the arithmetic early: latency, token budgets, cost per user
• You give real reviews and keep PRs small; for the lead role, you have led a team
• You work well with ambiguity and can move quickly without sacrificing quality
• You want to make a Blue Phone your daily phone

Nice to have

• Model Context Protocol (MCP), and OAuth flows for connected accounts
• Computer-use or screen-control agents
• Android development
• Experience reviewing AI-generated code, or working with coding agents

Why join now

• The assistant is the heart of Blue, and we are early in building what it can do
• You will set how we measure whether Blue is getting better