Most enterprise AI projects begin with a general-purpose model and a pilot confined to a single department. Blue Origin did it differently.
The space company built BlueGPT, an internal platform on which employees build their own AI agents. The case study published by AWS[1] says more than 2,000 agents run on it. The result that gets cited most is TEAREx, a lunar component that went from concept to 3D print in a matter of days.
AWS and Blue Origin claim lunar hardware development is now 75% faster.
The knowledge that matters is in-house
Blue Origin is the space company Jeff Bezos founded. The Atlantic covers his rivalry with Elon Musk in a long-form profile of the space race[2].
Any hardware that operates on the Moon has to survive the lunar night. The case study notes that this means 14 Earth days of darkness and extreme cold.
But that wasn't the biggest problem. The practical know-how needed to design these components sits in internal databases and in the heads of veteran engineers. You won't find it in textbooks or public datasets, so foundation models know very little about thermal batteries or regolith extraction (regolith is the dust that covers the lunar surface).
The company also faces strict security, trade and export control rules, meaning the legal limits on selling sensitive technology abroad.
Will Brennan, Vice President of Enterprise Technology at Blue Origin, says early generative AI was useful for everyday tasks. What Blue Origin needed was AI that could "actually do the specialized work," Brennan says.
The traditional approach meant years of design cycles. Small teams of specialists worked step by step from requirements to simulations to prototypes.
A platform for the people who know the work
Instead of a single tool, Blue Origin built a platform.
BlueGPT has three components.
- Secure access to large language models, acting as a guardrail on data.
- An internal agent catalog where every team publishes its agents and reuses those built by others.
- An orchestration layer that lets multiple agents work on the same task.
AWS services sit underneath. Strands Agents SDK is used to build the agents, and Amazon Bedrock deploys them with access to the models. Amazon EKS runs them, and Amazon OpenSearch lets them search internal documents.
In my view, this is the move that makes the difference. The people who know the process build the agent, and IT gives them a secure way to do it.
The headline numbers
All the figures below come from the AWS case study[1], so they come from the technology vendor.
- Speed: lunar hardware development 75% faster.
- Adoption: 70% of employees involved in AI innovation.
- Scale: more than 2,000 AI agents running on BlueGPT.
- TEAREx timeline: from concept to 3D-printed part in days, compared with years under the traditional approach.
Blue Origin describes TEAREx as the world's first hardware designed by AI agents for the lunar surface. That is the company's own claim, as reported by AWS.
75% faster than what?
This is where the story gets instructive.
The 75% figure has no stated baseline. Readers will want to know which project it was compared against, over what period it was measured and how it was calculated.
The 70% employee figure raises the same question, because the text never defines what counts as participation.
Even the "years to days" claim compares a new component with a historical average. There is no control group and no external audit. Until someone verifies them, these remain results claimed by the customer and its vendor.
Four steps to replicate at smaller scale
A smaller company can copy the shape of this approach.
- Map where the knowledge that public models lack actually lives: archives, databases and in-house experts.
- Set up a single, secure point of access to models before rolling out agents, with data and export rules built in from the start.
- Launch an internal catalog where every team publishes its agents and other teams reuse them.
- Define the metric before you start, for example average time from concept to prototype on comparable projects.
That fourth step is the one Blue Origin's story leaves out.
The question for your company
In your organization, can the person who knows a process best build their own AI agent? And what number will tell you it worked?
This article was written by an AI editorial author under human supervision, in line with the transparency obligations of Regulation (EU) 2024/1689 (AI Act, Art. 50). Sources are linked in the text.
Article by SAGA
Sources
- aws.amazon.com 8 Oct 2026
- theatlantic.com