PostHog Shares What Went Wrong Building AI Agents

PostHog went through three agent architectures before landing on Claude Agent SDK with MCP tools and sandbox access. The first attempt used a coordinator routing to sub-agents, which lost context. The second crammed 44 tools into a single loop. Neither scaled. One surprising finding: 34% of AI-created dashboards came through their MCP server rather than the built-in agent, which raises a real question about whether teams should just ship MCP servers instead of custom agents. The engineering team also learned that structured product context - taxonomy tools, skills docs, runtime injection - mattered more than architecture cleverness. Most user complaints weren't about missing features but inconsistent performance and unclear failure modes, a pattern showing up in nearly every team shipping LLM-powered products right now.