AWS pushes deeper into autonomous AI at re:Invent

AWS has used re:Invent 2025 to introduce a broad set of AI upgrades that stretch from agent infrastructure to new chips and model families. The announcements show a clear direction: more autonomy, more efficiency and fewer barriers between prototype and production.

Bedrock AgentCore strengthens push for purpose-built AI infrastructure

Companies working with agentic systems often struggle with long-running tasks, unpredictable workloads and the need for secure storage and retrieval of context. Many teams still build this foundation by hand, pulling together components and maintaining them over time.

AWS says Amazon Bedrock AgentCore offers a managed alternative. It supports frameworks such as CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK and Strands Agents, and handles core operational functions behind the scenes. Early adopters include organisations across sports, healthcare, telecom and software development. Downloads have already crossed two million in five months.

Customer momentum from early deployments

The press release highlights varied use cases:

• PGA TOUR has built a multi-agent content system that produces digital articles. Its teams now generate ten times more content with a 95 per cent reduction in costs.

• MongoDB used AgentCore to compress multi-week evaluation cycles into a single deployment path. By linking its AWS footprint with MongoDB Atlas as the Knowledge Base, the company delivered an agent-driven application in eight weeks rather than months.

• Swisscom standardised agent development using AgentCore Runtime, Identity and Memory. This helped the company move its consumer-facing agent to production in four weeks, focusing on personalised sales and technical support.

Model customisation becomes faster and more accessible

Running AI agents at scale is expensive, especially when models spend much of their time on routine steps. AWS is addressing this with new capabilities in Amazon Bedrock and Amazon SageMaker AI.

• Reinforcement Fine Tuning (RFT) in Bedrock raises accuracy by an average of 66 per cent compared with base models. Some customers, such as Salesforce, report improvements above 70 per cent.

• SageMaker AI serverless customisation reduces build and experimentation cycles from months to days. Collinear AI used it to shorten model iteration from weeks to days.

Kiro powers bring specialised knowledge to agents

Developers often need agents that understand specific tools such as dashboards, UI design systems, databases or APIs. Kiro powers add this domain knowledge in a single click. A power can include specialised tool access, steering files and action triggers.

Examples listed include powers for Datadog, Dynatrace, Figma, Neon, Netlify, Postman, Stripe, Supabase and AWS. Developers can also create and share their own powers, loading them only when needed to keep token usage efficient.

SageMaker HyperPod introduces checkpointless training

Large-scale training is vulnerable to hardware or software faults. Traditional checkpoint recovery can take up to an hour, which stalls clusters and adds cost. AWS is now offering checkpointless training on SageMaker HyperPod, which recovers in minutes without manual steps.

The platform maintains model state continuously across accelerators. When faults appear, the system replaces failing components and restores training using peer-to-peer state transfers. AWS says this can push cluster efficiency to 95 per cent for very large training jobs.

Strands Agents expands to TypeScript and edge devices

Strands Agents, initially released in Python, now comes to TypeScript in preview. Developers gain full type safety, async/await support and integration with the AWS CDK. Downloads of the Python version have already crossed three million.

AWS is also adding general availability for edge device support, allowing autonomous agents to run on small, local hardware. This targets areas such as automotive, gaming and robotics. Developers can run local models through tools like Ollama and Llama.cpp, with bi-directional streaming between devices and the Cloud.

Amazon expands Nova model family and introduces open training

AWS is adding four new models to its Nova family. These are aimed at reasoning, multimodal tasks, conversational systems, code generation and agentic workloads. The more notable addition is Nova Forge, which introduces the idea of “open training.” Customers can access pre-trained checkpoints and blend Amazon datasets with their own.

The release cites Nova Act’s 90 per cent reliability on browser-based UI automation. Reddit has replaced several specialised models with this system. Hertz used the Nova Act to raise development speed by five times.

Frontier agents mark a shift toward long-running autonomous systems

AWS is also rolling out three “frontier agents,” described as long-running and scalable:

• Kiro autonomous agent, positioned as a virtual developer.

• AWS Security Agent, operating as an internal security consultant.

• AWS DevOps Agent, built as an operational responder.

Early users include Commonwealth Bank of Australia, SmugMug and Western Governors University, each using at least one of the agents to change how software teams work day to day.

Trainium3 UltraServers increase compute performance for large-scale training

Training advanced models remains costly, with only a few organisations able to fund the infrastructure needed. AWS has introduced EC2 Trn3 UltraServers, powered by its first 3nm AI chip. Each system packs up to 144 Trainium3 chips, delivering:

• Up to 4.4× more compute performance than Trainium2 UltraServers.

• Four times the energy efficiency.

• Three times the throughput per chip.

• Four times faster response times.

Customers such as Anthropic, Karakuri, Metagenomi, NetoAI, Ricoh and Splash Music report cost reductions of up to 50 per cent. Decart is seeing four times faster inference for real-time generative video at half the GPU cost. Amazon Bedrock is already using Trainium3 for production workloads.

Closing view

The breadth of updates signals a clear shift in AWS’s AI roadmap. Infrastructure is becoming more modular. Training is becoming more resilient. Agents are moving from short-lived scripts to long-running digital workers. And model control is widening through open checkpoints and rapid customisation.

For developers and enterprises in India and the wider SAARC region, the message is straightforward: expect faster build cycles, deeper automation and wider distribution of agentic workloads across both Cloud and edge environments.

Read More:

India hit by 265 million cyberattacks: Seqrite’s 2026 report warns of escalating threats

CrowdStrike on empowering India’s channel partners for cybersecurity’s future

Quick Heal version 26: anti-fraud, dark-web monitoring and partner growth

Green IT in India: Why sustainable digital infrastructure Is no longer optional