Credit: Amazon

Amazon has announced a new wave of capabilities for Amazon Bedrock AgentCore, designed to remove long-standing barriers preventing AI agents from reaching production at scale. In just five months since preview, AgentCore has already surpassed 2 million SDK downloads and is powering enterprise-grade agentic systems across industries.

Major early adopters include:

• PGA TOUR, which built a multi-agent content generation system achieving 1,000% faster content creation and 95% cost reduction.

• Workday, which uses AgentCore Code Interpreter for secure financial intelligence—reducing planning analysis time by 30%.

• Grupo Elfa, relying on AgentCore Observability for complete audit traceability, improving resolution times by 50%.

With organizations scaling agent deployments, teams face challenges around guardrails, safety, data access, and quality. The latest release squarely addresses these needs—bringing policy enforcement, real-time evaluation, long-term learning, and natural conversational streaming to the platform.

Key New Capabilities

Policy in AgentCore (Preview): Guardrails with Fine-Grained Control

The new Policy system allows enterprises to define strict, enforceable boundaries for what agents can do—independent of model choice.

• Intercepts tool calls before execution

• Uses natural language or Cedar, an open-source policy language

• Supports validation, auditability, and compliance

• Works across APIs, Lambda functions, MCP servers, and external services

This ensures agents act autonomously within approved limits, giving organizations the confidence to deploy them widely.

AgentCore Evaluations (Preview): Real-Time Quality Intelligence

AgentCore Evaluations provides continuous monitoring of agent behavior using built-in quality metrics such as:

• Correctness

• Faithfulness to context

• Helpfulness

• Harmfulness & stereotyping

• Tool selection accuracy

Teams can also create custom evaluators tailored to business rules. All results are streamed into Amazon CloudWatch, enabling dashboards, alarms, and automated responses.

This allows enterprises to detect degraded behavior quickly and stop faulty versions before they reach production.

Episodic Memory in AgentCore: Experience-Based Learning

AgentCore Memory now supports an episodic long-term memory strategy, enabling agents to learn from real past interactions.

Agents can now:

• Recognize repeating user patterns

• Apply past lessons to new tasks

• Improve consistency and reduce task time

• Avoid bloated instructions by retrieving only relevant learnings

This makes agents behave more like real assistants—learning preferences, adapting decisions, and improving over time.

Bidirectional Streaming for Natural, Interruptible Conversations

AgentCore Runtime now supports true two-way voice streaming:

• Users can interrupt the agent mid-response

• Agents adjust context instantly

• Conversations flow naturally—no rigid turn-taking

• Infrastructure for simultaneous input/output is handled automatically

This unlocks next-generation voice assistants for support, service, sales, and internal tools.

Building Trusted AI Agents at Scale

By combining policy enforcement, behavioral evaluations, episodic memory, and natural conversation handling, Amazon Bedrock AgentCore provides the most complete agentic platform for enterprises today.

These capabilities give developers and security teams confidence that agents will:

• Operate safely

• Respect data boundaries

• Maintain output quality

• Improve over time

• Deliver human-like conversational experiences

All while running on scalable, production-ready Bedrock infrastructure.

Availability

The new Policy and Evaluations features are available in preview, while enhanced Memory and bidirectional streaming are available today through AgentCore gateways, runtime, and console.