Dr. Swami Sivasubramanian, Vice President of Agentic AI, made the case that AWS' suite of AI tools is best suited for wrangling AI agents and customizing models to deliver business outcomes.
Speaking at his AWS re:Invent 2025 keynote, Sivasubramanian said:
"The question isn't whether you should customize your models, but how quickly can you get started?"
The future to Sivasubramanian is custom quality models that can carry out enterprise-specific tasks efficiently. "As agents become easier to build, the next big question emerges, how do we make them more efficient? Today's off the shelf models have broad intelligence. They can handle complex to use, multi-step reasoning and unexpected situation, but they aren't always the most efficient," said Sivasubramanian. "And this efficiency is not just about cost. It's about latency. How quickly can your agent respond? It's about scale. Can it handle quick demand? It's about agility. Can you iterate and improve quickly?"
Sivasubramanian said the barrage of announcements from re:Invent 2025 were about removing complexity and costs for model customization without an army of PhDs. Sivasubramanian followed up on earlier re:Invent announcements revolving around Amazon AgentCore, AWS Marketplace and multiple other products.
More from re:Invent 2025
- AWS' Kiro launches autonomous agents for individual developers
- AWS adds AI agent policy, evaluation tools to Amazon Bedrock AgentCore
- AWS launches Amazon Nova Forge, Nova 2 Omni
- AWS launches AI factory service, Trainium 4 with Trainium 4 on deck
- AWS Transform aims for custom code, enterprise tech debt
- AWS, Google Cloud engineer interconnect between their clouds
- AWS Marketplace adds solutions-based offers, Agent Mode
Here's a look at the news items from Sivasubramanian's keynote:
- Amazon Bedrock is getting new model customization tools that features reinforcement fine-tuning models that can deliver accuracy gains of 66% over base models. Amazon Bedrock automates the reinforcement fine tuning workflows without needing machine learning expertise. Amazon Nova is the first model offered and with other models coming soon.
- Amazon SageMaker AI is gaining serverless customization for multiple AI models including Amazon Nova, DeepSeek, GPT-OSS, Llama and Qwen. Amazon SageMaker AI will support reinforcement learning via a simple interface. Models can be customized in days instead of months.
- Amazon SageMaker HyperPod will get checkpointless training, which streamlines checkpoint-restart cycles. Amazon SageMaker Hyperpod will also get elastic training to manage cluster utilization and prioritize workloads.
