© Carbon60 Operating Co LTD
© Carbon60 Operating Co LTD
If 2024 was the year of experimenting with AI assistants, 2025 is the year of AI Agents. That was the driving message behind AWS CEO Matt Garman’s keynote at re:Invent 2025.
Assistants respond to prompts.
Agents plan, act, and adapt. They write and ship code, analyze logs during outages, conduct market research, or orchestrate multi-hour workflows with minimal human input.
AWS unveiled a massive slate of infrastructure, models, and tools to make this “agentic” future a reality. Here is the complete breakdown of every announcement from the keynote
AWS introduced two platforms aimed at reshaping how teams work: one for developers and one for the general workforce.
A new spec-driven development environment where engineers describe the intent, and Kiro generates the implementation. Instead of typing code line by line, teams define outcomes, constraints, and architectural requirements, and Kiro generates working code.
Instead of wrestling with boilerplate, legacy frameworks, or brittle integrations, engineers become orchestrators of intent. This collapses cycle times, reduces cognitive load, and allows smaller teams to deliver work previously reserved for larger engineering groups.
With Amazon Quick, you get a consumer-grade, enterprise-ready AI application that connects to corporate data (Salesforce, Microsoft 365, ServiceNow, and more).
Quick Flows let non-technical users build low-code “mini-agents” for tasks like compliance reporting or multi-source data consolidation.
Deep Research is an autonomous research agent that can synthesize internal documents and web sources into fully cited reports.
Garman introduced a new class of long-running, multi-step agents — “Frontier Agents” — designed to run business-critical workflows with minimal supervision.
Kiro Autonomous Agent: Takes a backlog ticket (e.g., "update this library across 15 microservices"), plans the work, edits code across multiple repositories, runs tests, and opens a Pull Request without requiring any manual intervention.
AWS Security Agent: Embeds into the development lifecycle to review design docs and code for vulnerabilities before they ship. It can also perform on-demand penetration testing.
AWS DevOps Agent: An always-on operator that detects incidents, independently analyzes root causes (e.g., tracing an outage to a specific bad IAM policy change), and proposes fixes instantly, enabling faster incident response and improved reliability.
At the heart of the agentic AWS stack is the expanded Nova model family. Through the Amazon Nova 2 model family, AWS now offers a robust set of foundation models and tools to customize and deploy them at enterprise scale.
Nova 2 Omni: The industry's first unified multimodal model. It accepts text, audio, video, and image inputs and generates text or image outputs in a single pass (no stitching multiple models together). Omni unlocks new workflows like automated document parsing, video analysis, and cross-modal compliance scanning with far less complexity.”
Nova 2 Sonic: A speech-to-speech model with human-like latency and emotion, designed for next-gen customer service bots.
Nova 2 Lite & Pro: Updated reasoning models. "Lite" is a cost-effective workhorse; "Pro" handles complex agentic reasoning.
Beyond out-of-the-box models, AWS introduced:
Nova Forge: A new “open training” service that lets organizations blend their proprietary data into Nova during training, creating customized “Novella” models optimized for their domain.
Nova Act: A service for building and managing reliable AI agents that automate browser-based UI workflows, powered by a custom Nova 2 Lite model. Early customers reportedly see 90% reliability for UI automation tasks.
Together, these make Nova not just a model suite but a full stack for agentic AI, from foundation models to customization to deployment.
To support large-scale AI and agent workloads, AWS is expanding its custom silicon and offering new infrastructure deployment models.
Trainium 3 (Available Now): The first 3nm AI chip in the cloud. It offers 4.4x more compute and 5x better power efficiency than Trainium 2.
Trainium 4 (Preview): Already in design, promising another massive leap in performance.
Trainium3 UltraServers: New servers built on AWS’s 3 nm AI chips designed to deliver high compute power and energy efficiency for heavy model training and inference workloads. This move strengthens AWS’s ability to run frontier AI models at scale.
AWS AI Factories: AWS will now ship racks of their custom compute (Trainium and Inferentia) directly to your data center. It functions like a private AWS region behind your firewall, solving data sovereignty issues.
Ocelot Quantum Chip: A prototype chip that reduces quantum error correction costs by 90%, a major step toward commercially viable quantum computing.
New Nvidia Instances: G6 (Blackwell GB200) and G6 UltraServers (GB300) for massive model training. Enterprises looking to fine-tune or train agents now have best-in-class GPU options next to AWS’s custom silicon.
Building agents is easy; trusting them is hard. AWS introduced tools to solve hallucinations, policy violations, and unexpected agent behavior.
Bedrock Agent Core Policy: Allows you to set deterministic "hard rules" (e.g., "Never process a refund over $1,000 without human approval") that override the AI's decision-making.
Bedrock Agent Core Evaluations: An automated testing framework that grades your agents on helpfulness, safety, and brand alignment before they are deployed.
Lambda Durable Functions: Allows serverless functions to "pause" and wait for long periods (days or weeks)—critical for agent workflows that need to wait for human approval or external events without racking up costs.
AWS extended AWS Transform with a “Custom” mode capable of modernizing almost any legacy codebase — ERP languages, monoliths, VBA scripts, or aging front-ends.
Agent-driven modernization allows organizations to tackle years of tech debt in months, not years. The cloud migration bottleneck finally becomes manageable for mid-market and enterprise teams alike.
Garman delivered over twenty foundational updates across databases, storage, compute, and security. AI adoption surges only if the underlying cloud substrate scales effortlessly with it. AWS is quietly reinforcing the reliability, throughput, and cost efficiency needed to support a world where agents generate more logs, more events, more storage operations, and more compute consumption.
In short, AWS is upgrading the cloud to withstand the next decade of AI-driven traffic. Here’s a quick look at all the announcements.
Database Savings Plans: A unified savings plan offering up to 35% off across RDS, Aurora, DynamoDB, and more.
RDS Storage: Storage limits for SQL Server and Oracle increased from 64TB to 256TB.
SQL Server Updates: Support for Developer Edition (no licensing fees for dev/test) and vCPU optimization controls to lower licensing costs.
EMR Serverless: No longer requires provisioning local storage.
OpenSearch: GPU acceleration for vector indexing (10x faster).
S3 Object Size: Maximum file size increased from 5TB to 50TB.
S3 Tables: Now supports Intelligent-Tiering (auto-saving costs on unused data) and Cross-Region Replication.
S3 Vectors: Generally Available (GA).
FSx: S3 Access Points now support NetApp ONTAP.
X-Family Instances: New large-memory instances (Xeon 6) with 50% more memory.
AMD EPYC Memory Instances: Now up to 3TB of memory.
C8in Instances: (Phonetic: "C8 ion") Optimized for high-performance networking (2.5x packet performance).
M8zn Instances: (Phonetic: "Mate AZN") Fastest CPU clock frequency in the cloud for gaming and high-frequency trading.
Mac Instances: New M3 Ultra and M4 Max Mac instances.
GuardDuty: Extended Threat Detection now covers EKS (Containers) and EC2 at no additional cost.
Security Hub: Now GA with real-time risk analytics and a simplified pricing model.
CloudWatch: A new unified data store for all security and compliance logs.
The barrier to entry for intelligent automation has collapsed. Kiro removes the friction of coding; Nova Forge eliminates the friction of model customization; and Frontier Agents remove the friction of operations.
However, the complexity of implementation remains, from data readiness to ensuring security and scalable infrastructure. Agentic workflows aren’t just about turning on a service; they require a fundamental rethinking of your data strategy and operational foundations. Agents can only plan and act effectively if the underlying data they access is clean, accessible, and governed.
As an AWS Premier Partner, OpsGuru specializes in building the secure, scalable cloud foundations required to support autonomous agents. From data modernization to the deployment of custom Nova models, we help you get it right the first time. with our experts to explore how you can establish your agentic AI foundations on AWS today.
