Senior Software Engineer, Content Agent

Sanity; United States; Hybrid; USD 121,000 - 202,000; Full time

Job summary

Sanity is hiring a senior engineer for Content Agent, the ops layer for structured content. Work across Dashboards, Studio, and APIs to build MCP tools, orchestration, and prompts that power AI-driven content workflows.

You’ll shape how content operations scale with reliable agent behavior and system-level quality guarantees. Location is London, UK or Oslo, Norway with in-office 2–3 days per week; you’ll thrive shipping LLM-based systems in production and thinking through edge cases,

Qualifications

• Shipped LLM-based systems in production.
• Strong TypeScript and React across the stack.
• Experience designing reliable agent systems.
• Ability to reason about user intent and edge cases.

Responsibilities

• Full-stack product work: React interfaces in Dashboards, APIs, backend services.
• Develop MCP tools and integrations.
• Design agent orchestration logic and human escalation.
• Create system prompts as engineering artifacts.
• Run evals and experiments to improve reliability in real usage.
• Leverage GROQ for content understanding.

Skills

TypeScript

React

LLM systems

System design

Probabilistic systems

Tools

GROQ

Sanity API / MCP tools

Job description

We're hiring a senior engineer for Content Agent - Sanity's operations layer for structured content. It's the system that turns content work (bulk edits, audits, translation, publishing) into things an agent can actually run. It lives in Dashboard, Studio, Slack, and behind a headless API. If you've spent time on AI systems that have to hold up for real users, not just in a demo, read on.At Sanity, we build an AI-powered content operating system that lets teams model, create, and automate content the way their business actually runs. SKIMS, Figma, Riot Games, Anthropic, COMPLEX, Nordstrom, and Morningbrew run their content operations on it.Content Agent is the part of the product that understands structured content - schemas, references, field types, validation rules - and can be trusted to operate on it correctly. You'd be doing full-stack product work in TypeScript and React: MCP tools and integrations, agent orchestration, system prompts, evals to make the whole thing more reliable. Engineers on this team own their domain and work closely with product on what gets built next.

What you'd be doing:

• Full-stack product work: React interfaces in Dashboards, APIs, backend services.
• MCP tools and integrations that extend what the agent can do.
• Agent loops: the orchestration logic that decides what the agent does, in what order, and when to bring a human in.
• System prompts. We treat these as real engineering artifacts, not a note at the top of a file.
• Evals and experiments to measure and improve reliability in real usage.
• GROQ, Sanity's content query language - using it to give the agent a deep understanding of structured content.
• Shaping what Content Agent becomes, with product and design. There's a lot of latitude here.

About you:

• Location: London, UK or Oslo, Norway (in-office 2–3 days/week)
• Strong TypeScript and React, comfortable across the stack (frontend, APIs, backend). You don't want to be boxed into one layer.
• You've shipped LLM-based systems in production. You know the hard part isn't getting a model to do something once; it's getting it to do the right thing for every user who shows up.
• You think in systems. Given a new agent capability, you think through user intent, edge cases, failure modes, and what that means for the architecture: what the agent knows, what tools it has, when it escalates.
• You write well and you care about why one phrasing works where another doesn't. An agent's instructions are prose running on a system that reads literally. Writing them well takes a feel for how language lands, not just what it says - who "you" refers to, what's implied versus explicit, how a sentence will actually be parsed.
• You're comfortable with probabilistic systems. Improving an LLM product means running experiments, designing evals, reading results statistically, and not chasing individual failures when the aggregate trend is what matters. You've moved past "it compiles or it doesn't."
• Real experience with, or real curiosity about, MCP, tool-use patterns, an