Forward Deployed Engineer, Agent Runtimes & MCP (Python)

Forward Deployed Engineer, Agent Runtimes & MCP (Python)

Mission. Build the Agent Trust integration into Anthropic's Claude Commerce Agents blueprint and the surrounding Claude tooling and make it adoptable by merchant engineers in a day. What you will do in the first 90 days Stand up the blueprint's verticals on all three runtimes; run its test and verification suites; pin and maintain our fork compatibility. Implement the AgentTrustMixin for the blueprint's Storefront Backend / Merchant Backend: identity bound at session start, in-path trust checks on cart and checkout hooks, decision mapping to the blueprint's native tool-result semantics (Unavailable, domain errors), per-hook fail policy, decision logging without PII, outcome reporting. Build and test the storefront MCP server path for Managed Agents behind an authenticating gateway, including vault-credential and signed-claim identity options. Write an Experian MCP server (advisory tools) and a companion skill for the blueprint's commerce-builder Claude Code plugin. Author the merchant integration guide and measure the merchant effort (files, lines, prompt bytes, gates, time-to-first-verified-session) as an acceptance metric. Must have 5+ years Python; strong async (asyncio), typed code, pytest, ruff; comfortable subclassing abstract backends and mapping exceptions to structured results. Shipped tool-using LLM agents on the Claude Messages API or Agent SDK; understands prompt caching mechanics and why static-prefix bytes matter. Built or integrated MCP servers (streamable HTTP) and put authentication in front of them; understands that a model can skip a tool call and enforcement must live server-side. JWT/JWKS verification, OAuth client credentials, secrets handling; can call a real-time risk API inside a latency budget and reason about fail-open vs fail-closed. Reads an unfamiliar, unmaintained open-source codebase quickly and cites module and line in every claim; works from written acceptance criteria without supervision.