LoginSign Up
← Back to all sparks
A

AWS Machine Learning

AI-ASSISTANTS
Velocity10.0

Amazon Web Services' official AI/ML blog covering Bedrock, SageMaker, AgentCore, and Nova model updates.

aws.amazon.com/blogs/machine-learning ↗

AWS keeps building the agent operations layer, now with wallets and spending limits.

agentcorebedrockagent paymentsagent observabilitymodel catalog
◆Current state
The AWS ML feed is almost entirely Bedrock AgentCore at this point: observability, browser automation, payments, and multi-agent orchestration, each shipped as a reference architecture rather than a product announcement. SageMaker AI has been demoted to a model-hosting substrate that AgentCore calls into. Amazon Quick's Microsoft 365 extensions remain the only recent piece aimed at an end user rather than a platform team.
◆Where it's heading
AWS is competing on the operational surface around agents rather than on models themselves — identity, tracing, cost attribution, payment rails, and monitoring that reaches agents running on GCP, Azure, or a laptop. The newest posts extend that in two directions at once: outward to agent-initiated payments over x402, and inward to keeping the JumpStart model catalog current. The tutorial-heavy cadence suggests the primitives are considered stable and the work is now proving enterprise patterns on top of them.
◆Prediction
Expect agent payments to move from testnet walkthroughs to a generally available, policy-governed capability, with spending guardrails surfaced as a first-class AgentCore control alongside identity and observability.

◆Recent moves

  1. 1d ago

    NVIDIA Nemotron 3.5 Lightning now available in Amazon SageMaker JumpStart

    Another open model lands in the SageMaker JumpStart catalog, this time a 30B Mixture-of-Experts with 3B active parameters aimed at always-on agentic workloads. The throughput and task-completion claims fit the agent-serving theme, but catalog additions are routine maintenance on this feed rather than a change of direction.

    View source ↗
  2. 1d ago

    Build OpenClaw agents that transact with Amazon Bedrock AgentCore payments

    A walkthrough for giving an agent a wallet and spending guardrails so it can pay for paywalled APIs and MCP servers via the x402 protocol. Payments were already among the AgentCore primitives, so this is the demonstration layer catching up — bounded, human-approved, and still on testnet.

    View source ↗
  3. 4d ago

    Custom reward functions for multi-turn reinforcement learning with Amazon Nova Forge

    Guidance on designing composite reward functions for multi-turn reinforcement learning on Nova Forge, including running model-generated code safely inside the reward. Useful craft detail, but no capability changed hands.

    View source ↗
  4. 4d ago

    Building agentic workflows with SageMaker AI and Bedrock AgentCore

    A pattern for routing each agent in a workflow to the model best suited to it, pairing OpenAI-compatible SageMaker endpoints with the AgentCore runtime. The token-observability gap it patches in Strands Agents is the only concrete detail; the rest is assembly instructions for existing parts.

    View source ↗
  5. 5d ago

    Monitor on-premises and multi-cloud AI agents with AgentCore Observability

    AgentCore Observability is documented for agents running outside AWS entirely — on-premises, on GCP, on Azure, or on developer machines — routed through the AWS Distro for OpenTelemetry. Extending the monitoring plane past the AWS boundary is the clearest statement yet that AWS wants to own agent operations regardless of where the agent runs.

    View source ↗
  6. 5d ago

    Automate legacy web applications with Amazon Bedrock AgentCore Browser Tool

    A reference architecture for driving legacy web interfaces through isolated AgentCore browser sessions with audit trails and human oversight. It restates the Browser Tool's existing pitch against a specific enterprise scenario.

    View source ↗