How Coinbase Builds Developer Support Agents | Interrupt 26

Evan Kormos, engineering manager at Coinbase, shares how a small developer support team went from zero to a self-improving, customer-facing agentic system for crypto developers. The talk covers three live agents: Discord AI Chat (with deterministic guardrails and LLM-as-judge response evaluation), a Slack Triage channel that surfaces Discord signals internally with direct links to LangSmith traces, and a Support Engineer Assistant for internal case management. Evan describes the team's intent-driven workflow — using Claude Code with MCP tools for tech grooming — and their signal search process for back-testing traces to find improvement opportunities. He closes with three lessons: treat agent engineering as its own discipline, build the glass box before the agent, and remember that the team is the multiplier. Chapters: 0:00 Introduction and what the Coinbase developer platform is 1:16 The support challenge: Discord, AI chat, and scaling for new markets 2:37 System design: delivery channels, Python services, and self-hosted LangSmith 4:49 Shared tools: remote MCP server for docs and RAG fallback 5:34 Overview of agents shipped: Discord AI Chat, Slack Triage, Support Engineer Assistant 6:22 Discord AI Chat demo: expert developer responses with guardrails 7:24 How guardrails work: deterministic safeguards plus LLM-as-judge output review 8:02 Slack Triage: surfacing Discord signals internally with LangSmith trace links 8:38 Support Engineer Assistant: internal case management surface 9:31 Intent-based workflow: Claude Code + MCP tools for tech grooming 10:37 Signal search: back-testing traces to find improvement opportunities 11:22 What we found: multilingual adversaries in production 12:08 Three takeaways: agent engineering as a discipline, glass box first, team as multiplier Resources: → LangGraph: https://www.langchain.com/langgraph → LangSmith: https://www.langchain.com/langsmith → LangChain Academy: https://academy.langchain.com