Applied AI Engineer - Agent Harness

Summary

Join the team redefining what a deeply personal and integrated assistant can be. As part of the Siri organization, you will help shape one of the world’s most widely used AI assistants. It runs on our next generation of Apple Intelligence, with capabilities like personal context understanding and on-screen awareness, and it is built with privacy from the ground up. We build the harness, the foundation that turns a model into an assistant that can act. This is a systems engineering role for people who understand how large language models behave.

Minimum Qualifications

Bachelor's degree in Computer Science or a related field, with 2+ years of industry experience Strong programming skills in a systems language such as Swift, C++, Rust, Objective-C, or Go Solid foundation in concurrency, inter-process communication, and performance optimization Hands-on experience building LLM-powered systems such as agents, tool calling, or prompt and context design, through professional work or substantial personal projects Daily use of agentic coding tools such as Claude Code, Codex, or Pi, with informed opinions on what makes their harnesses succeed or fail Ability to communicate clearly and make progress on ambiguous problems

Preferred Qualifications

Experience with Swift and Apple platform development Experience building agent harnesses, orchestration frameworks, or multi-step tool-use systems Experience with model inference integration, streaming protocols, or latency-sensitive client-server systems Experience using evals to diagnose and improve LLM behavior Experience designing APIs or platforms adopted by other engineering teams Familiarity with privacy and security engineering in production systems Active engagement with current research and practice in agent harness design, such as context engineering, tool use, and long-horizon agents, with a record of turning new ideas into practical improvements

Description

A model can't act on its own. The harness runs the agent loop, carries context and tools to the model, executes the tool calls it makes, and streams results back reliably. It spans models running on device and larger models in Private Cloud Compute. It handles everything from single-turn requests to long multi-step tasks that span apps, survive interruptions, and resume cleanly. We're looking for a strong systems engineer who is fluent in how LLMs behave. The work is real-time and resource-constrained: responses must start fast enough for a voice assistant to feel instant, under tight latency, memory, and power limits on everything from Mac and iPhone to Apple Watch. Failures rarely have a single cause. In a typical week you might track down why a streamed response stalls, make tool calls cancel cleanly when the user interrupts, or replay a failed request to tell whether the model, the context, or the runtime was at fault. Every change to the harness changes how the model behaves, so we measure our work with evals, not just tests. You'll join the team that owns the core runtime and work alongside senior engineers and researchers. You'll also partner closely with the modeling, tools, and context teams and with app and framework teams across Apple. You'll take features from design through production and grow your ownership of the system over time. You'll use agent harnesses every day while building one for an assistant used by hundreds of millions of people.

Responsibilities

- Build and evolve core parts of the agent runtime: the planning loop, session lifecycle, concurrency, and cancellation - Build the transcript and conversation state, and the platform that defines tools and dynamic instructions and delivers them to the model - Build the executor that dispatches tool calls into Apple's apps and services and handles their results and failures reliably - Integrate on-device and server models, including low-latency streaming and interruption mid-response - Build safety, security, and privacy protections into harness components, working with security and privacy partners - Validate harness changes with evals and latency metrics, and adapt the harness as each new model generation arrives - Build tracing, replay, and testing tools that make a nondeterministic system debuggable