AWS Weekly Roundup — February 2026: AgentCore, Bedrock, EC2 and More

February 2026 has been one of those months where you blink and there are fifteen new blog posts on the AWS News Blog. The throughline? AI agents are no longer a demo — they’re an infrastructure category. AgentCore went from “interesting preview” to “here’s how you build production systems with it” in the span of two weeks. Meanwhile, EC2 keeps getting new silicon, and the European Sovereign Cloud is officially open for business.

Here’s my curated take on what matters and why.

🤖 AgentCore & AI Agents — The Main Event

If you’re building anything with AI agents on AWS, this was your month. Six posts, all substantive, all pointing in the same direction: AgentCore is becoming the opinionated runtime for agentic workloads.

CAKE: Unified Intelligence with AgentCore (Feb 18)

AWS published a deep dive on CAKE — Customer Agent & Knowledge Engine — built entirely on AgentCore. The architecture is what caught my eye: Neptune for graph relationships, DynamoDB for transactional state, OpenSearch for semantic retrieval, and web search — all orchestrated in parallel via Strands Agents.

This isn’t a toy demo. They’re doing multi-tool coordination with row-level security, meaning the agent can only access data the authenticated user is permitted to see. If you’ve been wondering how to build a real customer-facing agent that queries multiple backends without turning into a security nightmare, this is your reference architecture.

Why this matters: Most agent demos hand-wave the hard parts — auth, data isolation, concurrent tool execution. CAKE tackles them head-on. Study this pattern.

Long-Running MCP Servers on AgentCore (Feb 12)

This post addresses a real pain point: what happens when your agent task takes hours, not seconds? The answer involves context messaging for heartbeats (so the caller knows the agent is still alive), async task management for fire-and-forget patterns, and persistent state across invocations.

These are the production patterns that separate a hackathon project from a system you’d actually deploy. If you’ve tried building agents that do multi-step research, document processing, or data pipeline orchestration, you know the pain of long-running execution. This post gives you the primitives.

Why this matters: Agents that can only do 30-second tasks are glorified function calls. Long-running execution is what makes them genuinely useful.

AgentCore Browser: Proxies, Profiles, Extensions (Feb 13)

Three new capabilities for agent browsing sessions: proxy routing (critical for corporate networks behind firewalls), persistent browser profiles (cookies, sessions, login state preserved across invocations), and Chrome extensions running inside agent browser sessions.

The extensions piece is particularly interesting — think ad blockers, authentication extensions, or custom DOM manipulation tools running inside an agent’s browsing context. This is getting close to “agent as a power user” territory.

Why this matters: Browser-based agents that can’t handle proxies or persist sessions are useless in enterprise environments. These are table-stakes features being shipped.

FAST: Full-Stack Starter Template (Feb 9)

AWS released FAST (awslabs/fullstack-solution-template-for-agentcore) — a CDK template that wires up AgentCore Runtime, Gateway, Memory, Code Interpreter, a React frontend, and Cognito auth. It supports both Strands and LangGraph agent frameworks.

One cdk deploy and you have a working agent application with auth, memory, and a UI. This is the kind of accelerator that should have existed six months ago.

Why this matters: Reduces time-to-first-agent from days to hours. If you’re evaluating AgentCore for a project, start here.

Evaluating AI Agents (Feb 18)

Perhaps the most underappreciated post this month. AWS shared their internal framework for evaluating agents — and it goes well beyond “did the agent get the right answer.” They evaluate reasoning quality, tool selection accuracy, error recovery, and multi-step plan coherence.

AgentCore Evaluations bakes this into the platform. If you’ve been eyeballing agent quality with vibes, this is the structured alternative.

Why this matters: You can’t improve what you can’t measure. Agent evaluation is the missing piece for most teams, and this is a credible approach.

Quick Agents + MCP Tool Integration (Feb 20)

A more accessible entry point: integrating external tools into Quick Agents via the Model Context Protocol (MCP). MCP continues to gain traction as the standard interface between agents and tools, and this post shows the practical wiring.

Why this matters: MCP is becoming the USB-C of agent tooling. Learning it now pays dividends across frameworks.

☁️ Infrastructure Updates

New EC2 Instance Families

Three launches worth noting:

• HPC8a (Feb 16) — Powered by AMD EPYC 5th gen (Turin). Purpose-built for tightly-coupled HPC workloads. If you’re running computational fluid dynamics or weather modeling, these are your new best friend.
• G7e with NVIDIA Blackwell GPUs (Jan 20) — RTX PRO 6000 (Blackwell) on EC2. Inference, rendering, and simulation workloads get a significant perf/watt improvement.
• C8id, M8id, R8id (Feb 4) — The “id” suffix means local NVMe, and these go up to 22.8 TB. Compute-optimized, general-purpose, and memory-optimized flavors. Great for caching layers, temporary storage-heavy workloads, and databases that benefit from local disk.

My take: The steady cadence of new instance types is easy to tune out, but the NVMe-dense instances (C8id family) solve real problems for anyone running Elasticsearch, Redis with persistence, or large temporary datasets.

AWS European Sovereign Cloud (Jan 15)

It’s official — the AWS European Sovereign Cloud is now open. Operated by EU-resident staff, data stays in the EU, and it’s designed to meet the strictest European data residency and sovereignty requirements.

Why this matters: If you work with EU public sector, financial services, or healthcare customers, this changes the conversation from “we can’t use cloud” to “which workloads go first.”

IAM Identity Center Multi-Region Replication (Feb 3)

Your IAM Identity Center configuration can now replicate across regions. This is a quiet but important improvement for disaster recovery and multi-region architectures. Previously, Identity Center was a single-region resource — a gap that made multi-region designs awkward.

🧠 AI/ML Beyond Agents

SageMaker + Nova Custom Models (Feb 16)

You can now deploy custom-trained Amazon Nova models via SageMaker Inference. This bridges the gap between Bedrock’s fine-tuning capabilities and SageMaker’s flexible deployment options (custom containers, autoscaling, multi-model endpoints).

Bedrock Reinforcement Fine-Tuning

Reinforcement fine-tuning in Bedrock lets you go beyond supervised fine-tuning with reward-model-guided optimization. This is the RLHF loop, but managed. If you’ve been doing SFT and hitting a quality ceiling, this is the next lever to pull.

NVIDIA Nemotron 3 Nano 30B MoE on JumpStart (Feb 11)

A 30B Mixture-of-Experts model from NVIDIA, available one-click on SageMaker JumpStart. MoE architectures give you large-model quality with smaller-model inference costs (only a subset of parameters activate per token). Worth benchmarking against your current model choices.

Nova Sonic for Voice Assistants (Feb 10)

Amazon Nova Sonic enables real-time, streaming voice interactions. Think conversational AI with low latency — IVR replacements, voice-controlled agent interfaces, accessibility tools. The streaming aspect is key: no more waiting for full transcription before responding.

Claude Opus 4.6 in Amazon Bedrock (Feb 9)

Anthropic’s latest flagship model is now available in Bedrock. Opus 4.6 brings improved reasoning, longer context handling, and better instruction following. If you’re using Claude via Bedrock, this is a straightforward upgrade — same API, better model.

🔭 Key Takeaways

1. AgentCore is the platform bet. Six substantive posts in two weeks isn’t a coincidence — AWS is positioning AgentCore as the default way to build, deploy, and evaluate AI agents. If you haven’t explored it yet, the FAST template is your on-ramp.

2. MCP is winning. Between AgentCore’s native support and the broader ecosystem adoption, Model Context Protocol is becoming the standard for agent-tool communication. Invest in learning it.

3. Evaluation is the new frontier. Building agents is getting easier. Knowing whether they’re actually good is still hard. The evaluation framework post deserves a careful read.

4. Sovereignty is shipping. The European Sovereign Cloud moves from announcement to reality. Expect this to unlock workloads that were previously cloud-resistant.

5. The silicon treadmill continues. Blackwell GPUs, AMD Turin, 22 TB of local NVMe — the compute menu keeps expanding. Review your instance choices quarterly.

---

This roundup was curated with the help of an AI assistant monitoring AWS RSS feeds, blog scraping, and changelog diffs. Because if you’re writing about AI agents, you might as well use one to help track the news. See you next month.