Agentic AI

Amazon Cognito is now a core skill in the Agent Toolkit for AWS, enabling AI coding agents to configure user pools, app clients, and OAuth flows. The aws-auth skill supports token management, JWT authorizers, and passkey enrollment, allowing agents to implement secure sign-in patterns autonomously. When used with the AWS MCP Server, agents execute CLI commands with IAM guardrails and CloudTrail logging for auditability.

• AI agents can now autonomously set up and troubleshoot Cognito auth flows via the aws-auth skill.
• Covers user pools, app clients, OAuth 2.0, JWT authorizers, and WebAuthn enrollment.
• Agents can leverage IAM guardrails and CloudTrail audit logging when paired with AWS MCP Server.
• Works standalone via AWS CLI for direct command-line execution without full agent integration.
• Accelerates implementation of secure sign-in for users, AI agents, and microservices.

Google has open-sourced a repository of Agent Skills designed to streamline interactions with Google Cloud and related technologies. The project provides pre-built configurations for tasks such as authentication, solution architecture, and agentic analytics. Users can install specific skills via npx, allowing for modular adoption of these tools into their existing AI agent workflows.

• Enables modular installation of GCP-specific agent capabilities via npx.
• Covers key workflows like onboarding, authentication, and solution architecture.
• Supports agentic analytics across multi-cloud and data lakehouse environments.
• Repository is under active development, so features may evolve rapidly.

Amazon Bedrock AgentCore now generally offers runtime instances that deploy agents directly onto customer-owned EC2 infrastructure. This new capability complements the existing microVM option by providing access to diverse instance types, including GPU, memory-optimized, and compute-optimized families. AWS manages the provisioning, patching, scaling, and lifecycle of these resources while teams define needs via capacity providers.

• Use EC2 runtime instances for sustained, resource-intensive, or specialized-hardware AI agent workloads.
• Access broad EC2 families including GPU, memory, and compute-optimized types directly within AgentCore.
• AWS handles infrastructure lifecycle management: provisioning, patching, scaling, and lifecycle ops.
• Configure capacity providers via Console, CLI, SDKs, or APIs to specify required instance types.

Amazon Bedrock AgentCore introduces temporal policies to enforce stateful authorization rules that evaluate requests based on prior agent actions within a session. This allows for workflow sequencing, argument validation against previous outputs, human approval gates, and data freshness checks. Additionally, new rate limiting capabilities enable per-user or per-group traffic controls for tools, models, and agents scoped via OAuth or IAM.

• Enforce workflow sequencing by evaluating requests against prior session actions.
• Require human approval before executing privileged agent actions.
• Validate tool arguments against outputs from previous calls in the session.
• Apply rate limits per user or group using OAuth or IAM scopes.
• Prevent stateless tool calls that are safe in isolation but risky in context.

Amazon Bedrock AgentCore has expanded to AWS GovCloud (US-West) with new capabilities for building regulated AI agents. The update introduces memory features for both short-term conversation context and long-term persistent insights. It also includes centralized policy controls for agent-tool interactions that operate independently of agent code.

• AgentCore is now available in AWS GovCloud (US-West) for regulated environments.
• Short-term memory handles immediate conversation context without manual infrastructure management.
• Long-term memory extracts persistent insights and user preferences across sessions.
• Centralized policies enforce fine-grained controls on agent-tool interactions outside code.
• These features aim to accelerate the transition from prototype to production for enterprise agents.