AI Platform Engineer, Agent Systems

Description

The company is building the agent platform for professional music production: the orchestration layer, tool interfaces, skills runtime, and context architecture that allow any AI agent to reason about and act on a music-production workflow.

You will lead the design of the orchestration loop, define how the engine’s capabilities are exposed to models, build the skills runtime that transforms a general-purpose model into a domain specialist, and architect the context and memory systems that keep agents coherent across long creative sessions.

The object model is a song. The users are producers, musicians, and creatives. The domain has real-time constraints, deep semantics, and no existing playbook.

Requirements

• Five or more years shipping production platform or infrastructure software that other engineers have built on top of.
• Eighteen or more months of production experience building LLM agent systems, covering orchestration loops, tool use, and context management. We have no preference for a specific framework. We are equally interested in engineers who shipped on a provider-agnostic framework such as LangGraph and engineers who rejected frameworks entirely and built their own harness, provided you can articulate what you learned from the path you took.
• Demonstrated experience designing tool interfaces for LLM consumption. You can explain what makes a tool schema discoverable and usable by a model versus merely technically correct.
• Demonstrated experience building context, memory, or state-management systems beyond framework defaults, including compaction, durable memory, or session persistence. You have diagnosed agent failures from raw execution traces and made targeted harness changes in response.
• Strong proficiency in TypeScript and Python.
• Experience with the Model Context Protocol (MCP) or similar tool-connectivity standards.

Nice to have (not required)

• Background in music production, audio engineering, or another creative-tool domain, including as a serious hobbyist.
• Experience with real-time audio systems, professional audio software, or other latency-sensitive environments.
• Experience making a complex desktop or professional application agent-accessible, in any domain with a rich object model (DAW, IDE, design tool, CAD).
• Experience building middleware or hook architectures that allow others to customize agent behavior without modifying core code.

Job responsibilities

What You Will Own

• Tool interfaces. Define how the engine’s capabilities are exposed to LLMs as structured, discoverable tools. This includes schemas, semantic descriptions, scoped tool sets, input validation, and output parsing that a model can reliably produce and the harness can reliably consume. Designing a tool surface that models use well is a distinct discipline from designing an API for human developers, and you will own that discipline.
• Orchestration and control flow. Design and build the harness: the core loop and the machinery around it. This covers step sequencing, retries, timeouts, error recovery, fallback paths, and multi-agent coordination where a workflow is split across sub-agents with their own tools and context. You will evaluate whether to build this in-house, adopt a framework, or extend an existing one. We have no commitment to any specific framework, and we will not build the platform on top of a single provider or model. A well-reasoned argument for building our own harness is a welcome outcome of that evaluation.
• Skills runtime. Design the format, packaging, loading, and execution layer for the structured domain knowledge that turns a generic model into a music-production specialist. This is our most distinctive platform primitive and it is largely greenfield.
• Context, memory, and state. Build the systems that keep agents performant and coherent across long, multi-step creative workflows. This includes context compaction, short-term working memory, durable cross-session memory, session state persistence, continuity across disconnects, and sub-agent delegation in which parent and child contexts remain consistent.
• Extension points. Design the harness so that new tools, skills, and middleware can be added without modifying the core runtime. Extensibility is an architectural property of the system, not a retrofit.
• Evaluations, observability, and failure analysis. Evaluations tell us the harness is working; raw execution traces and structured failure logs tell us why it is not. You will build and own the platform-level evaluation surface, the observability that every engineer on the platform depends on, and the feedback loop that converts failed agent runs into targeted harness changes.
• Ongoing simplification. As frontier models improve, some of the scaffolding we build today will stop earning its keep. You will audit the harness on a regular basis and remove the components that models no longer require.

This Role Is Not

• LLM integration engineering. This role is not responsible for wiring models to the DAW or building end-user AI features. This role builds the platform those features run on.
• ML or model engineering. This team does not train models. It builds the systems that agents run on.
• Research. This team applies current research in production. Original research happens elsewhere in the company.