AWS announced their new Agent Core platform at re:Invent, designed to help enterprises deploy AI agents with enhanced trust and control through integrated policy management, evaluations, and episodic memory. These features enable scalable, secure, and continuously improving agentic systems compatible with any AI model, addressing the high failure rate of AI pilots in production.
The video discusses recent major announcements made by AWS at their annual re:Invent conference, focusing on their agentic platform called Agent Core. This platform aims to address two critical challenges enterprises face when deploying AI agents in production: trust and control. The speaker references a viral MIT report revealing that 95% of AI pilots in enterprises fail, highlighting the urgency of solving these issues. AWS’s Agent Core stands out by integrating policy handling, evaluations, and episodic memory as core features, enabling enterprises to build, deploy, and scale agentic systems without the need for infrastructure management, and compatible with any AI model or framework.
One of the key new features introduced is policy management, which acts as guardrails to ensure agents only perform authorized actions. The platform allows users to define policies in natural language, which are then automatically converted into programmatic code. This makes it easy to set up and test policies that control agent behavior at scale, processing thousands of requests per second. The policy engine intercepts agent requests, verifies permissions, and restricts access to tools and data accordingly. This approach is crucial for enterprise deployments, especially given concerns about AI models’ potential for deception, lying, or unauthorized data access.
Another significant update is the introduction of evaluations within Agent Core. Evaluations provide a systematic way to measure and improve agent performance, which is often overlooked but essential for production-grade AI systems. Users can leverage pre-built evaluation templates or create custom ones to assess various aspects of agent behavior, such as correctness, helpfulness, coherence, and instruction-following. Continuous or on-demand evaluations offer full observability, allowing enterprises to trace errors or false outputs back to their root causes, thereby enhancing reliability and trustworthiness in deployed agents.
The third major enhancement is the addition of episodic memory to Agent Core. This feature enables agents to learn from past successes and failures across multiple interactions, recognizing patterns and applying those learnings to future tasks. Unlike memory tied to individual users or conversations, episodic memory spans the entire agent implementation, improving overall agent intelligence and adaptability. This memory system works hand-in-hand with evaluations, allowing the platform to measure whether agents are genuinely improving over time based on their accumulated experiences.
In summary, AWS’s Agent Core platform now incorporates policy management, evaluations, and episodic memory as foundational elements, all integrated deeply into the agent execution path rather than as afterthoughts. These innovations address the enterprise challenges of trust, control, and continuous improvement in AI agent deployment. The speaker encourages viewers to explore Agent Core, highlighting its potential to solve the enterprise rollout dilemma for AI agents. The video is sponsored by AWS, and links to resources and tutorials are provided for those interested in building and experimenting with agentic systems on this platform.