TLDR OpenClaw + an agent framework + Gemini is a practical default production architecture for long-context, tool-heavy agents. Gemini 3.1 Pro is OpenClaw's recommended default model ( google/gemini-3.1-pro-preview ), mostly for its 1M token context window and native tool-use support. LangChain/LangGraph is one of the most widely used combinations for Gemini + OpenClaw workflows in production. Frameworks covered: LangChain, LangGraph, CrewAI, AutoGen, Google ADK, SmolAgents. A code-review agent fetches a pull request, analyzes three files, flags an issue, then loops back to re-examine the diff. Somewhere between steps two and three, tool call history drops out of context. The agent hallucinates a fix or stalls. The reasoning step succeeded. The state didn't persist. These failures are rarely dramatic. They show up as lost context, repeated tool calls, inconsistent state between steps. OpenClaw runs execution and manages tool runtimes. Agent frameworks control orchestration and workflow logic.…