Staff+ Software Engineer, Claude Managed Agents

Anthropic•San Francisco, CA

1d•$405,000 - $485,000•Hybrid

About The Position

We are looking for experienced backend and distributed systems engineers to join the Agentic Systems team within our Platform organization. Agentic Systems builds Claude Managed Agents: the hosted platform for building, running, and scaling production agents on Claude. Instead of every developer hand-rolling an agent loop, sandboxed execution, state management, credential handling, and error recovery — and reworking all of it with every model release — Managed Agents pairs an Anthropic-built agent harness with production infrastructure for sessions, environments, tools, memory, and permissions, exposed through a small set of composable APIs designed to stay stable as models and harnesses evolve. It powers agentic products inside Anthropic as well as those built by customers on the Claude Platform. Manaaged Agents is in public beta and growing quickly, and this is still an early team with a lot of surface area left to define. You'll drive 0 → 1 efforts from ideation through GA, own systems end to end from API design through operations, and partner closely with product, research, developer experience, and go-to-market teams to figure out what "managed" should mean for the next generation of agents. You should be comfortable going deep on hard distributed systems problems, care about APIs as a product in their own right, and be motivated by turning ambiguous ideas into high-quality, shipped platform capabilities that other engineers build their products on.

Requirements

Have a minimum of 8 years of practical experience as a backend, distributed systems, or infrastructure engineer
Have built and operated stateful, long-running, or high-throughput systems in production — workflow orchestration, streaming, storage, container or job orchestration — and can reason rigorously about durability, consistency, failure modes, and cost
Have strong product sense and treat API design as a craft; you care about the developer on the other side of the interface and can ideate and execute product strategy with cross-functional partners in new domains
Are excited by 0 → 1 work and comfortable navigating ambiguity, and have ideally operated in both early-stage and more mature team or company settings
Use Claude or other AI tools as a core part of how you build software, and have opinions about what makes an agent harness good
Take full ownership of your work — from design through build, deployment, and operations (including on-call), to iterating on and improving what you ship
Care about building systems that other engineers and businesses love to use, and about doing so safely

Nice To Haves

Built or contributed to an agent harness, agent framework, or LLM orchestration layer — tool execution, context management, memory, or multi-agent coordination
Worked on an AI or ML platform (model serving, inference infrastructure, developer tooling) at an AI lab or an AI-native product company, or led adoption of AI-driven development inside an engineering organization
Built evaluation or benchmarking infrastructure for LLM or agent systems
Experience with durable execution or workflow engines, sandboxed code execution, or container runtimes
Shipped public developer platforms, APIs, or SDKs used by external developers at scale

Responsibilities

Design and operate the systems underneath that: durable session and event storage, sandbox orchestration, streaming, scheduling, and multi-tenant isolation. Reliability, latency, and cost efficiency are product features here, and you'll own them in production.
Work alongside research to revisit harness assumptions with each model generation, build the eval infrastructure that measures harness quality against research baselines and real customer workloads, and hold the bar that lets us say our harness gets the most out of Claude.
Ship the capabilities that raise the ceiling on what those agents can do: outcome-driven execution where developers specify success criteria and a budget and Claude iterates until it gets there, multi-agent orchestration, memory, and the observability and tracing that make long-running agents debuggable.
Shape the primitives — versioning, ergonomics across API, SDK, and CLI, sensible defaults, escape hatches — with the expectation that the implementations underneath will change many times while the contracts hold.