AWS re:Invent 2025 Highlights: AI Agents Take Center Stage in Vegas
AWS held its annual re:Invent conference in Las Vegas, drawing tens of thousands for keynotes from CEO Matt Garman and other announcements. The focus landed on autonomous AI agents, new chips, and tools for enterprise AI. Here’s what stood out across reports from The Tech Portal, TechBuzz, The Indian Express, and The Register.
Autonomous AI Agents for Long-Running Tasks
AWS pushed hard on “agentic” AI—systems that handle multi-step actions without constant oversight. Garman highlighted agents that run for hours or days. TechBuzz named three “Frontier agents”: Kiro for coding like a virtual developer, AWS Security Agent for code reviews, and AWS DevOps Agent for incident prevention.
The backbone is Amazon Bedrock AgentCore, a platform for building and deploying agents at scale. New features include Policy (in preview) for natural-language boundaries on actions, Evaluations with 13 pre-built checks for correctness and safety, and episodic Memory so agents learn from past runs, per The Indian Express. The Tech Portal noted Amazon CEO Andy Jassy praising AgentCore on X for improving security and scalability, with building blocks like Policy and Evaluations added recently.
Nova Forge pairs with AgentCore to manage multi-agent workflows. Early adopters in finance and healthcare use these for custom data integration while keeping audits, according to AWS engineers cited by The Tech Portal.
New AI Models on Bedrock
Bedrock expanded with Nova family models: multimodal Nova 2 Omni (text, speech, images, video) in preview, plus Nova 2 Lite and Nova 2 Sonic. Nova Forge lets customers build custom models for better accuracy. The platform now has about 100 serverless models, including 18 new open-weight ones—four from Mistral AI available first on Bedrock—as reported by The Indian Express.
Hardware: Trainium3 Chips and Graviton5 Processors
AWS targeted Nvidia with Trainium3 accelerators for training multimodal and long-context models. EC2 Trn3 UltraServers powered by Trainium3 deliver 4.4x compute performance, 4x energy efficiency, and 4x memory bandwidth over predecessors, according to The Indian Express. TechBuzz cited 4x gains in training and inference, plus 40% lower energy use. These power new EC2 UltraServer clusters.
AWS previewed Trainium4 with at least 6x processing performance and 3x FP8 performance. Jassy noted on X that Trainium2 already brings in multi-billion-dollar revenue.
Separately, Graviton5 packs 192 Arm Neoverse V3 cores on TSMC 3nm, with 192MB L3 cache and memory speeds up to 7200 MT/s (8800 planned). M9g instances hit 25% higher performance than Graviton4-based M8g ones, using a single socket to cut latency. Graviton chips power over half of new AWS CPU capacity for three years running. It supports PCIe 6 and suits gaming, databases, EDA, and analytics, The Register, Investing.com, and About Amazon reported.
AI Factories and Other Tools
AWS AI Factories let enterprises run models in their data centers with Trainium3 or Nvidia GPUs, plus Bedrock and SageMaker. Customers supply space and power; AWS handles setup to cut timelines from years to months. The Tech Portal, TechBuzz, and The Indian Express all covered strong interest from governments and regulated sectors.
Other updates: VMware migration tools to ease cloud shifts, serverless AI agents for coding and call centers (The Tech Portal), and Bedrock tweaks for enterprise tuning.
