AWS AI Updates: November 7, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Moonshot AI’s Kimi K3.1 Pro Arrives on Amazon Bedrock as a Highspeed Open-Weight Option
Amazon Bedrock added Kimi K3.1 Pro (Highspeed) from Moonshot AI on November 6, served at a high output-tokens-per-second rate aimed at agentic coding and document processing. Moonshot describes the model as open-weight, natively multimodal, with a 1-million-token context window and its most capable 3-trillion-parameter-class model so far. On Bedrock it runs inside the same security boundary as the proprietary models and supports explicit prompt caching to cut latency and input cost, and it is offered in every Region where Bedrock is available through cross-Region inference. AWS did not publish pricing in the announcement. Source
2. SageMaker Inference Adds NVIDIA MIG Partitions, and a Half-H100 Slice Beat a Whole A100 by 53% to 89% on Throughput
Amazon SageMaker Inference now supports GPU partitioning with NVIDIA Multi-Instance GPU, so one accelerator can be split into up to seven hardware-isolated partitions, each with its own compute, memory and memory bandwidth, and a different model can run in each. It works only on endpoints that use inference components: an AcceleratorPartitionLayouts field on the production variant in CreateEndpointConfig declares how every GPU is carved (for example mig-4g.20gb, mig-2g.10gb and mig-1g.5gb on each A100 of an ml.p4d.24xlarge, giving 24 partitions), and an AcceleratorPartitionType field on each component picks its slice; there are no new API operations or IAM actions, and carving added about 17 seconds to provisioning. Supported families are P4d (A100), P5 (H100), G7 (RTX PRO 4500 Blackwell, two partitions) and G7e (RTX PRO 6000 Blackwell, four partitions) in seven Regions, and the layout is fixed at endpoint creation, so changing it means a new endpoint. In AWS’s genai-bench test of Llama-3.1-8B-Instruct, one mig-3g.40gb partition of an H100 delivered 53% to 89% more output tokens per second than a whole A100-40GB at the load where its cost advantage peaked, with 5.6% to 47.5% lower time to first token and 18% to 34% lower cost per million output tokens above a crossover (concurrency 64 for 100-token prompts, 4 for 8,000-token prompts); below that crossover the whole A100 was cheaper. Salesforce helped define the requirements and validated the first implementation. Source
3. AWS Sample Runs Claude Code as a Remote Agent in One AgentCore microVM per Task, Returning a Pull Request
AWS published sample-remote-coding-agent, which runs Claude Code as a remote coding agent on Amazon Bedrock AgentCore so a task can be handed off and returned later as a GitHub pull request. A Lambda function screens each task prompt with a Bedrock Guardrails guardrail and queues it through DynamoDB and SQS, and each registered repository gets its own AgentCore runtime whose sessions run in separate microVMs carrying the Claude Code CLI, the Claude Agent SDK, git and make. Inside a session the agent takes a short-lived GitHub App token from AgentCore Identity, checks AgentCore memory for lessons from earlier tasks on the same repository, plans with Claude Opus and writes code with Claude Sonnet on Bedrock, runs the repository’s own setup and test targets, and opens a pull request that stays a draft unless verification passed; it never merges. The runtime role can invoke only Anthropic models on Bedrock through a private endpoint, outbound traffic goes through a NAT gateway with a DNS Firewall that blocks the vendor’s own domains such as api.anthropic.com (AWS recommends a Network Firewall allowlist for production), and deployment takes 15 to 25 minutes with idle charges and a default $200 monthly Bedrock budget alert that does not cap spending. Source
4. AWS and Okta Show Agent Delegation Through Cross App Access, With Revoked Users Refused in About a Second
AWS published a two-part guide to governing agents that act for users by combining Okta Cross App Access (XAA) with AgentCore. Okta decides which registered agent may act for which assigned user against which resource: per request, the backend exchanges the user’s ID token for a single-use Identity Assertion JWT Authorization Grant (ID-JAG) that expires in 5 minutes, AgentCore runtime’s JWT authorizer accepts only grants Okta issued to that agent, and AgentCore Identity redeems the grant for a short-lived access token kept in its token vault and never exposed to agent code. In AWS’s tests, Okta refused new requests for a user about 1 second after an administrator removed that user’s assignment, already issued tokens expired within 5 minutes, and AgentCore Identity added about 0.6 seconds to a 30-second request. Part 2 moves the pattern behind AgentCore Gateway as a single front door that redeems Okta’s grant, binds each tool to an agent with Cedar policies, adds a second agent on Amazon EKS and logs every allow and deny in CloudWatch next to Okta’s system log. Source Source
5. Greenlight Financial’s Luna SOC Agent Resolves More Than 90% of Up to 4,000 Monthly Cases Without a Human
Greenlight Financial described Luna, a tiered multi-agent security operations system on AgentCore runtime and memory that triages dark web breach-card alerts, identity anomalies, infrastructure alerts and financial-risk events that take a human analyst 15 to 30 minutes each. A Tier 1 supervisor fed by API Gateway (with an MCP endpoint) and Tines SOAR orchestrates more than 30 specialized agents and settles routine cases in single-digit seconds, while roughly 10% of cases escalate to a Tier 2 supervisor on a reasoning model; engineers can also drive investigations from a Claude Desktop extension. A 30-day episodic memory plus long-term learnings (for example allowlisting a recurring maintenance access pattern after one resolution) lifted autonomous resolution from 83% with Tier 1 alone to more than 90%, which Greenlight estimates saves 900 to 1,800 analyst hours a month. Luna triages and recommends but does not remediate without explicit policy approval. Source
6. Amazon Quick Team Publishes Its Agent Taxonomy: Triggers, Agents, Capabilities and Learned Context
AWS’s Amazon Quick team shared the vocabulary it settled on after its agents moved from answering questions to taking actions, built around four questions: When, What, Who and How. Triggers carry both the moment of activation and the instructions, and come in four kinds (human-initiated, agent-delegated, scheduled and event-driven), with the trigger type shaping behavior, such as a scheduled task doing nothing when unsure while a human-initiated one asks. An agent is defined as a persistent, learning actor made of standing instructions, dynamically updated capabilities (tools, skills, workflows and sub-agents) and learned context that accumulates through use. It is the first post in a series that will cover Quick’s agentic operating system, background learning and proactive feeds. Source
7. EC2 Capacity Blocks for ML Upfront Fees Are Now Amortized in CUR 2.0 and Cost Explorer
AWS now spreads the upfront fee for EC2 Capacity Blocks for ML across each reservation’s duration in Cost and Usage Report 2.0 and Cost Explorer, instead of showing a single charge on the purchase date. Teams can see the amortized cost of each instance launched into a block and of the capacity that went unused, which lets them charge reserved GPU spend back to the workloads and teams that used it. In CUR 2.0 the figures appear in the reservation_amortized_upfront_fee_for_billing_period and reservation_effective_cost columns, with reservation_net_effective_cost after discounts. The change is free in all Regions except GovCloud (US) and China. Source
8. AWS Elemental MediaConvert Generates Subtitles Inside Transcoding Jobs With Smart Subtitles
AWS Elemental MediaConvert now integrates Smart Subtitles, built on AWS Elemental Inference and recent speech recognition models, so video-on-demand jobs can transcribe audio into subtitles without a separate captioning pipeline. It supports English, French, German, Italian, Portuguese and Spanish, outputs side-car text formats or burned-in captions for HLS and DASH delivery, and accepts custom dictionaries in Elemental Inference for proper nouns, brand names and technical terms. It is available wherever Elemental Inference runs: US East (N. Virginia), US West (Oregon), Europe (Ireland), Asia Pacific (Mumbai) and Asia Pacific (Sydney). Source