Build long-horizon agents with Grok 4.7 on Amazon Bedrock

AWS published a walkthrough for building a tool-calling agent on Grok 4.7 in Amazon Bedrock and measuring how reasoning effort affects accuracy and token use. Across 72 runs on three search-style tasks, the agent answered 67 correctly, with all 18 runs at xhigh correct and 15 of 18 at low using about an eighth of the output tokens. Grok 4.7 is accessed through the us.xai.grok-4.7 or global.xai.grok-4.7 cross-Region inference profile, supports four effort levels (low, medium, high, xhigh), tool calling and structured output, and SpaceXAI lists a 500K token context window.