Blue Origin's In-Space Systems team used BlueGPT, the company's internal multi-agent AI platform, to design the Thermal Energy Advanced Regolith Exchanger (TEAREx) — a thermal battery system meant to help lunar hardware survive the 14-Earth-day lunar night — and compressed the hardware development cycle by roughly 90%, going from concept to printed part in days rather than years.
"The early days of GenAI were exciting for us, but limited to more common sorts of tasks," said Will Brennan, Blue Origin's VP of Enterprise Technology. "We needed AI that could actually do the specialized work, not just assist with it." Building lunar hardware that survives two weeks of freezing darkness pulls on deep specialization across thermal battery design, regolith extraction, and industrial manufacturing — knowledge that lives in internal systems, expert engineers' heads, and proprietary datasets, not in any base large language model, and is subject to strict security, trade, and export controls.
Blue Origin responded by building BlueGPT, a secure LLM gateway combined with an agent marketplace and a multi-agent orchestration platform. The stack uses the open-source Strands Agents SDK for model-driven agentic design; Amazon Bedrock — including Bedrock Knowledge Bases for proprietary RAG and Bedrock AgentCore for hierarchical short-term and long-term agentic memory — for foundation-model access and agent deployment; Amazon EKS for the agent runtime; Amazon OpenSearch for retrieval; Amazon RDS for general-purpose data; and AWS Lambda to scale serverless workflow automation. The platform is open to every Blue Origin employee and now hosts more than 2,700 agents that drove roughly 3.5 million interactions in the month prior to publication, with 70% company-wide adoption.
To design TEAREx, the In-Space Systems team assembled a coordinated cast of specialized agents — supervisors, librarians, requirements agents, design agents, and analysis agents — to work alongside a small group of two or three human engineers. Connected to AI-native physics-simulation tools running on GPU-accelerated Amazon EC2 P5 and G5 instances, the agents ran iterative design loops autonomously: generating candidates, evaluating them against requirements, iterating, and continuing until specifications were met. Manufacturing teams use the same agents to improve work orders and resolve non-conformances roughly 70% faster, engineers use coding agents to write code, and operations teams communicate design changes with suppliers through AI intermediaries.
The agents created for TEAREx are not one-off artifacts. They live in BlueGPT's marketplace, where they can be reused and recombined across tubes, barrels, harnesses, assemblies, and eventually entire vehicles — a pattern the company is now extending to the New Glenn orbital rocket and the Blue Moon lunar lander programs. The same autonomous workflow that compressed hardware development by 90% also accelerated analysis tasks by 6x (work that took four days now takes four hours), delivered mass-optimized designs through automated topology optimization, and let small teams of two or three engineers deliver the work of dozens.
"This is what we see as the future of engineering teams at Blue Origin," Brennan noted, describing small teams working with large teams of AI agents to deliver the work of dozens at orders-of-magnitude faster speeds. The company isn't stopping at TEAREx: it is building an enterprise knowledge graph on AWS to transform AI accessibility across all of its data, testing custom AI models with autonomous lunar rovers, and expanding agentic workflows across New Glenn and the lunar lander programs — laying the groundwork, in Brennan's framing, for "the road to space faster than ever before — for the benefit of Earth."