Build long-horizon agents with Grok 4.7 on Amazon Bedrock

來源原標題

Build long-horizon agents with Grok 4.7 on Amazon Bedrock

作者

Suheel Farooq

原始標籤

Advanced (300) · Amazon Bedrock · Amazon Machine Learning · Artificial Intelligence · Foundation models · Generative AI · Technical How-to

原始連結

https://aws.amazon.com/blogs/machine-learning/build-long-horizon-agents-with-grok-4-7-on-amazon-bedrock/

摘要原文

SpaceXAI’s Grok 4.7 is now available on Amazon Bedrock, which offers a choice of foundation models (FMs) from multiple providers. According to SpaceXAI, Grok 4.7 is trained to work longer on more difficult tasks and knows when to check its own work. That matters most for agents, which carry a task through many steps and must verify what they find along the way. An agent that works this way has to do more than look things up. It has to form a candidate answer, check it against what it has learned, and refine it until it fits. Grok 4.7 exposes a reasoning effort setting that controls how much deliberation happens before each response, and for agent workloads a practical question is which level to run. In this post, we build a tool-calling agent on Grok 4.7 and evaluate it on three tasks that require search rather than lookup. The tasks are diagnosing a two-service outage from its alerts, scheduling maintenance under interacting constraints, and choosing an optimal set of migrations under a budget. Each task has exactly one correct answer, verified by brute force. We ran all three at every reasoning effort level, six times each, for 72 agent runs. The agent answered 67 of the 72 runs correctly. At xhigh it answered all 18 correctly, and at low it answered 15 of 18 while using about an eighth of the output tokens. That gives a clear trade-off you can tune per request, and we show how to measure it on your own tasks and how to monitor it in production. We report token counts rather than currency throughout. Token prices change, and the counts are what your own costs are derived from, so a measurement expressed in tokens stays valid. According to SpaceXAI, Grok 4.7 uses a larger base model than Grok 4.6 and was trained with a longer reinforcement learning run on a harder mix of tasks. This mix is weighted toward problems that take many hours to complete. SpaceXAI reports that the model is better at verifying its own work and at managing longer context. SpaceXAI’s published benchmark results follow that emphasis. The largest reported gains over Grok 4.6 appear in the categories that measure sustained work rather than single-turn answers. Source: SpaceXAI, Introducing Grok 4.7 ( https://x.ai/news/grok-4-7 ). Scores are as published by SpaceXAI at the time of the model launch. Benchmarks show how a model performs across standard tasks. For your own workload, the useful next step is to measure on tasks shaped like yours, which is what the rest of this post walks through. SpaceXAI lists a 500K token context window for Grok 4.7, with text and image input and text output. On Amazon Bedrock, the model supports configurable reasoning effort, tool calling, and structured output. Grok 4.7 runs on the bedrock-runtime endpoint and is accessed through a cross-Region inference profile. Two profiles are available: Cross-Region inference routes requests across multiple AWS Regions, which gives you higher throughput without managing capacity Region by Region. Requests name one of the inference profiles rather than a bare model ID. The model supports the OpenAI-compatible Chat Completions and Responses APIs, and the Amazon Bedrock Converse API. On Converse, reasoning is returned as a reasoningContent block alongside the text block. In inferenceConfig , set maxTokens only, and put model-specific parameters, such as reasoning effort, in additionalModelRequestFields . For the current list of supported Regions, service tiers, and token prices, see the Grok 4.7 model card in the Amazon Bedrock User Guide and the Amazon Bedrock pricing page ( https://aws.amazon.com/bedrock/pricing/ ). Your AWS Identity and Access Management (IAM) identity needs bedrock:InvokeModel on both the inference profile and your account’s default project: Include both resources in the policy statement. The default project permission is required in addition to the inference profile. For production use, we recommend generating a short-term bearer token from your existing AWS credentials, so authentication stays tied to your IAM identity and no long-lived secret is stored: Reasoning tokens count as output tokens and draw down max_completion_tokens before the answer text is produced, so size the limit to leave room for both. If the limit is reached during reasoning, the response returns without answer text, so check for empty content in agent code and raise the limit if you see it. Our runs used a limit of 32,768 tokens, which left headroom at every effort level. Grok 4.7 supports four effort levels: low , medium , high , and xhigh . Effort levels can change between model versions, so it is worth confirming them against the service rather than hardcoding a list: Many agent evaluations test diligence: read a record, filter it, add it up. Those tasks are worth having, but to see what reasoning effort contributes, a task has to require search: the answer can’t be read off the data, it has to be proposed and then checked. We built a synthetic environment of 20 services, exposed through seven tools, and three tasks on top of it. Every task has exactly one correct answer, and each is built so that a quick heuristic is not enough to find it. Four practices keep a measurement like this trustworthy, and they apply to most agent evaluations. State every rule the grader relies on. If the prompt leaves an edge case open, a reasonable answer can be scored as wrong, and the evaluation ends up measuring the specification rather than the model. The diagnosis prompt, for example, says explicitly that a failed service raises no alert of its own. Prove each task has one answer. Each environment is small enough to enumerate, so we checked every candidate answer before running anything. For the schedule, that is a few lines: Grade a required final line, not the whole response. Each prompt asks the agent to end with a line such as FINAL: auth, checkout . The grader reads only that line, so a response that discusses other candidates while explaining its reasoning is graded on its conclusion: Track format separately from correctness. A missing final line is a different signal from an incorrect one. In our runs, all 72 responses included a gradeable final line. Constrain tool parameters with an enum where the valid values are known, so every tool call targets a real resource: The system prompt is deliberately short, and asks the agent to check its answer before giving it: The agent loop is a standard tool-calling loop: call the model with reasoning_effort set, execute each returned tool call, append the result keyed by tool_call_id , and repeat until the model stops requesting tools. Tracking prompt_tokens and completion_tokens across the loop gives the token counts reported in the following section. Six runs of each task at each effort level, 72 runs in total. The agent answered 67 correctly. Figure 1: Share of runs answered correctly at each reasoning effort level (left), and mean output tokens per run with median seconds per run overlaid (right) xhigh answered every run correctly. All 18 runs across the three tasks were correct at xhigh , and medium and high answered 17 of 18 each. With six runs per task and level, the differences between levels are indicative rather than statistically conclusive, and the direction was consistent across all three tasks. Effort changes how much the agent reasons, not how much it explores. Tool calls were essentially flat, between 17.2 and 18.7 per run at every level. The agent gathered the same facts regardless of effort, and what changed was how much reasoning it applied to them, with output tokens rising from 1,184 at low to 9,147 at xhigh . low offers a strong cost-to-quality ratio. At low , the agent answered 15 of 18 runs correctly using about an eighth of the output tokens of xhigh , with a median run of 18 seconds. Where a slightly lower completion rate is acceptable, or where answers are reviewed downstream, low is an efficient choice.