Amazon Bedrock AgentCore Adds Policy, Evaluations and Episodic Memory for Production AI Agents

Amazon Web Services has expanded its Bedrock AgentCore platform with three capabilities aimed at making AI agents safe to run in production: Policy, Evaluations and Memory. The additions address a gap that has slowed enterprise adoption of autonomous agents — the absence of observable, enforceable controls that sit outside the agent itself.

Policy: deterministic guardrails outside agent code

AgentCore Policy enforces boundaries on what an agent can do without requiring those rules to be embedded in the agent's own prompt or code. The controls run through AgentCore Gateway, the managed layer that brokers tool calls and API access. Because the policy layer is external to the model, it cannot be overridden by prompt injection or by the agent reasoning its way around a soft instruction.

AWS says Policy supports allow-lists and deny-lists for tool use, rate limits on specific actions, and scope restrictions that prevent an agent from accessing resources outside a defined boundary. Enterprises can define policies at the organisation, application or individual agent level, and changes take effect without redeploying the agent.

Evaluations: continuous quality and safety monitoring in production

AgentCore Evaluations extends the existing Bedrock model-evaluation tooling to cover live agent runs rather than only pre-deployment test sets. AWS ships pre-built evaluation dimensions covering response quality, task completion, safety and groundedness, and allows teams to define custom dimensions against their own criteria.

The service samples agent interactions continuously and surfaces results in a dashboard, with alerting when scores fall below configured thresholds. AWS positions this as closing the loop between pre-release testing and what agents actually do once deployed at scale — a distinction that matters because agent behaviour can drift as the tools and data sources they call change over time.

Memory: episodic learning from prior outcomes

The Memory capability adds episodic storage to AgentCore, allowing agents to retrieve structured records of past interactions and their outcomes. Unlike semantic memory, which stores facts, episodic memory stores sequences of actions and results, giving an agent the ability to recognise that a particular approach succeeded or failed in a comparable prior situation.

AWS says Memory is scoped per user and per application, with retention policies configurable by the operator. The company notes that episodic memory is distinct from the context window: it persists across sessions and is retrieved selectively rather than loaded in full, keeping inference costs predictable.