Amazon Web Services (AWS) has introduced three new artificial intelligence (AI) agents and expanded its AgentCore platform with features aimed at helping enterprises deploy and supervise AI systems more safely and consistently.
The announcements were made during AWS re:Invent 2025, the company’s annual cloud and AI conference held in Las Vegas.
New frontier agents for coding, security, and operations
AWS’s new agents are positioned as autonomous teammates across the software development lifecycle:
Kiro autonomous agent
Focuses on software development. It maintains persistent context across sessions, learns from a team’s repositories and workflows, and can work on complex tasks such as multi-repository code changes, bug triage, and test coverage improvements. AWS says Kiro can run independently on development tasks for hours or days with minimal human intervention.
AWS Security Agent
Embeds security checks into development. It reviews design documents and pull requests against organisational security standards, performs on-demand penetration testing, and returns validated findings with remediation code that development teams can apply.
AWS DevOps Agent
Supports incident response and operational reliability. It connects to observability tools, runbooks, and CI/CD pipelines to correlate telemetry and deployment data, helping teams find root causes faster and recommend improvements. Inside Amazon, AWS reports that the agent has already handled thousands of escalations and achieved an estimated root cause identification rate above 86 percent.
All three frontier agents are available in preview.
AgentCore adds policy, evaluations, and episodic memory
Alongside the new agents, AWS announced new capabilities in Amazon Bedrock AgentCore, its platform for building and deploying AI agents.
Key additions include:
Policy in AgentCore, which lets teams define natural-language rules that control what actions agents can take and what data or tools they may access. These rules are converted into Cedar policies and enforced via AgentCore Gateway, which intercepts tool calls in real time to block actions that violate configured constraints.
AgentCore Evaluations, a suite of 13 pre-built evaluators that score agents on dimensions such as correctness, safety, tool selection accuracy, goal success rate, and context relevance. Development teams can also write custom evaluators and use the service to monitor agents in testing and production.
AgentCore Memory, which introduces episodic memory so agents can learn from previous interactions. Each “episode” captures context, reasoning, actions, and outcomes, allowing agents to draw on prior experience when handling similar tasks in future.
AWS positions these features as foundations for “trusted AI agents,” giving enterprises more explicit control over autonomy, data access, and performance as they scale agentic workloads.
Relevance for Singapore and Asia-Pacific enterprises
The updates come as companies in Singapore accelerate AI adoption across banking, aviation, logistics, and the public sector.
Many of these industries face strict compliance and audit requirements, making governance tools a key factor in selecting AI platforms.
By integrating boundaries, policy controls, and evaluation systems directly into AgentCore, AWS aims to support organisations that need to deploy semi-autonomous workflows while maintaining operational and regulatory safeguards.
AWS positions AgentCore as a long-term foundation
AWS said it plans to continue expanding AgentCore’s capabilities, particularly in reasoning, safe automation, and complex multi-step workflows.
The company emphasised that enterprises should expect rapid iteration across agent memory, evaluation, and governance features as AI adoption increases globally.
