AWS AI Updates: April 23, 2026

1. Bedrock AgentCore Adds a Managed Agent Harness, a CLI, and Prebuilt Skills

AWS. AgentCore gained three new pieces aimed at cutting the boilerplate out of agent development. The managed harness (preview) lets teams define an agent by model, system prompt, and tool list without writing orchestration code — each session gets its own microVM filesystem and shell, and agents can switch models mid-session with filesystem state persisted across pause and resume. The AgentCore CLI deploys prototypes with infrastructure-as-code governance via AWS CDK, with Terraform coming. Pre-built skills optimize coding assistants, starting with Kiro Power and with support for Claude Code, Codex, and Cursor on the roadmap. Source

2. SageMaker AI Adds Serverless Fine-Tuning for Qwen3.5

AWS. SageMaker AI now supports serverless fine-tuning for Qwen3.5 at 4B, 9B, and 27B active parameters, covering both supervised fine-tuning and reinforcement fine-tuning. Customers supply the dataset and configuration; SageMaker handles cluster provisioning, training orchestration, and consumption-based billing. The move brings Alibaba’s Qwen3.5 family into AWS’s managed training story without requiring users to stand up their own GPU clusters. Source

3. SageMaker Unified Studio Supports Multiple Code Spaces Per Project

AWS. SageMaker Unified Studio now lets data workers in IAM-based domains spin up multiple JupyterLab and Code Editor spaces within a single project, each with its own EBS volume and runtime configuration. The change removes the one-space-per-project limit that forced teams to either serialize experiments or fragment work across projects. Spaces can open in separate browser tabs or connect to local IDEs, with Amazon Q paid-tier functionality available in each. Source

4. SageMaker AI Adds Inference Recommendations Backed by NVIDIA AIPerf

AWS. SageMaker AI’s new inference recommendations feature replaces the manual benchmarking grind for proprietary model deployment. Customers upload a model, describe anticipated traffic, and pick an optimization target (cost, latency, or throughput); SageMaker runs targeted configuration sweeps on actual GPU infrastructure via NVIDIA AIPerf and returns deployment-ready configurations with time-to-first-token, inter-token latency, request latency percentiles, throughput, and projected cost. The workflow is aimed at teams deploying custom fine-tuned or open-weight models who don’t want to hand-tune instance/config combinations. Source

5. Five Qwen Models Land in SageMaker JumpStart

AWS. JumpStart added Qwen3-Coder-Next, Qwen3-30B-A3B, Qwen3-30B-A3B-Thinking-2507, Qwen3-Coder-30B-A3B-Instruct, and Qwen3.5-4B. The lineup targets distinct niches: Qwen3-Coder-Next for long-horizon coding agents with heavy tool use; Qwen3-30B-A3B for mode-switching between fast and thinking responses; Qwen3-30B-A3B-Thinking-2507 for reasoning-heavy tasks; Qwen3-Coder-30B-A3B-Instruct for agentic coding with custom function calling; and Qwen3.5-4B as a lightweight multimodal option trained on 201 languages. Source