Build long-horizon agents with Grok 4.7 on Amazon Bedrock
Grok 4.7 is now available on Amazon Bedrock. In this post, we build a tool-calling agent on Grok 4.7 and evaluate it on three reasoning tasks at four reasoning effort levels, measuring correctness and token cost per completed task. We show how to tune the trade-off per request and monitor it in production.
SpaceXAI has made its Grok 4.7 model available on Amazon Bedrock, a platform that offers foundation models (FMs) from various providers. Grok 4.7 is designed to tackle complex tasks by working for longer durations and having the ability to self-check its own work. This feature is particularly important for agents, which are responsible for carrying out tasks that involve multiple steps and require verification of the findings.
An agent utilizing Grok 4.7 not only looks up information but also formulates a tentative answer, checks it against its existing knowledge, and refines it until it meets the desired criteria. The model provides a reasoning effort setting that allows users to control the amount of deliberation before each response. For agent workloads, the key question is determining the appropriate level of reasoning effort.
In this report, we build a tool-calling agent using Grok 4.7 and evaluate its performance on three tasks that require searching rather than direct lookup. These tasks include diagnosing a two-service outage based on alerts, scheduling maintenance under interacting constraints, and selecting an optimal set of migrations within a given budget.
Each task has a single correct answer, which was verified through brute force. We executed all three tasks at every reasoning effort level, repeating each run six times, for a total of 72 agent runs.
Out of these 72 runs, Grok 4.7 achieved 67 correct answers. At the highest reasoning effort level (xhigh), it answered all 18 tasks correctly, while at the lowest level (low), it answered 15 out of 18 tasks correctly while using approximately one-eighth of the output tokens. This demonstrates a clear trade-off between accuracy and token usage that can be adjusted based on the specific requirements of each request.
We measure this trade-off in tokens, which are the unit used to calculate costs, ensuring the measurement remains valid even if token prices fluctuate.
Written by urgent.news from AWS Machine Learning's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
