Build long-horizon agents with Grok 4.7 on Amazon Bedrock
Amazon Web Services Suheel Farooq
Grok 4.7 is on Amazon Bedrock now, and AWS tested it on agents. It nailed all 18 high-effort runs, but low used about one-eighth the output tokens.
Based on reporting by Amazon Web Services, Suheel Farooq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon Bedrock now offers Grok 4.7 from SpaceXAI, and AWS did something more useful than a launch-post victory lap: it measured the model on agent work that actually needs judgment. The focus here is long-horizon tasks, not chatty one-shot answers. Grok 4.7 is meant to work longer on harder problems, and SpaceXAI says it’s better at checking its own work and handling longer context. That is the part that matters when an agent has to search, compare, and verify before it speaks.
AWS built a tool-calling agent around Grok 4.7 and ran it through three synthetic tasks. One was a two-service outage where the failed services never raise their own alerts. One was a maintenance schedule with eight services, four nights, ten constraints, and a rule about team ownership. The last was a migration selection problem with a budget and prerequisites. Each task had exactly one correct answer, and each was brute-force verified before the model ever saw it.
The results draw a clean line between effort and cost. Across 72 runs, the agent got 67 right. At xhigh, it got all 18 runs right. At low, it got 15 of 18 while using about an eighth of the output tokens. Mean output tokens rose from 1,184 at low to 9,147 at xhigh, while tool calls stayed roughly flat. In other words, the model wasn’t exploring more. It was thinking harder about the same facts.
That distinction is the point. For the diagnosis and migration tasks, medium and above were enough to land the correct answer every time in this setup. The scheduling task was the stubborn one, and xhigh was the only level that solved all six runs. So the practical advice is simple: use lower effort when a small miss rate is acceptable, and spend the extra tokens when the answer has to be right the first time.
AWS also lays out the plumbing. Grok 4.7 runs through cross-Region inference profiles on bedrock-runtime, with us.xai.grok-4.7 for US data residency and global.xai.grok-4.7 for broader commercial AWS Regions. The model supports the OpenAI-compatible Chat Completions and Responses APIs plus Bedrock Converse, and reasoning shows up in a reasoningContent block. AWS also reminds builders to leave room for reasoning tokens in max completion limits, because if you cut it too close, you can get reasoning with no answer text at all.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI launch write-up that respects the budget line instead of treating it like decorative paperwork. The real story is not that xhigh won; it’s that low was already good enough to be useful, which is exactly how sane software gets adopted. The industry could use fewer victory laps and more tables like this one, especially from vendors who like to talk about “agents” as if they were magic interns.
Read more about this at: Amazon Web Services
