AWS unveils AI agents for security and DevOps, but builds guardrails to keep them in check

AWS pushes agentic AI into enterprise operations with new autonomous tools

AWS introduced a suite of AI agents at its New York Summit designed to operate continuously across security, DevOps, and workplace productivity 1

1

2

2

. The centerpiece announcements include AWS Continuum, a security agent now in closed preview that performs vulnerability scans and generates fixes, and significant updates to Amazon Quick, the workplace AI assistant that now lets users build autonomous agents using plain language 4

4

. Matt Wood, chief AI and technology officer, emphasized that these agentic AI systems should run continuously in the background rather than on demand, marking a shift from chat-based assistants to agents that complete tasks without constant human intervention 1

1

.

AWS Continuum tackles security vulnerabilities with supervised autonomy

AWS Continuum represents the company's answer to accelerating cyberattacks, particularly those enabled by advanced AI models like Anthropic's Claude Mythos that can chain minor flaws into critical exploits 2

2

. The security agent prioritizes findings that are actually reachable in production paths, demonstrates exploits in a sandbox environment, and generates suggested fixes including network changes or code patches 1

1

. Neha Rungta, AWS director of applied science who led Continuum's development, explained that AI can now combine two medium-severity findings and a low one into something critical, lowering the barrier for attackers 2

2

. Continuum starts in a supervised "learn mode" and earns the right to act independently only as customers grant permission category by category, revealing AWS's cautious approach to AI autonomy and human control 2

2

.

DevOps Agent gains release management as AWS addresses AI-generated code risks

The AWS DevOps Agent, made generally available in March after its preview at re:Invent in late 2025, now includes release management capabilities that assess code readiness and run software in AWS-managed isolated environments 1

1

. This addition addresses a growing problem: AI-driven DevOps tools now generate code at extraordinary speed while human review remains slow 4

4

. The agent supports Model Context Protocol (MCP) for calling tools and now exposes its own MCP endpoint, plus support for Google's Agent2Agent (A2A) protocol to enable agent collaboration 1

1

. DevOps Agent integrates with observability tools including AWS CloudWatch, Datadog, Dynatrace, New Relic, and Splunk, as well as repositories like GitHub and GitLab, and can connect to Microsoft Azure and Azure DevOps 1

1

.

Amazon Quick evolves from search assistant to autonomous agent platform

Swami Sivasubramanian, VP of Agentic AI at AWS, criticized first-generation AI assistants as "slightly faster search bars" that answer questions and forget, arguing the industry needs agents that create "compounding momentum" where each completed task feeds the next 3

3

. Amazon Quick now allows users to create background agents via voice prompts or choose from pre-configured agents, with a redesigned activity feed that triages email, messages, and calendar items into one prioritized view 1

1

2

2

. The service now connects to hundreds of third-party services including Gmail, Slack, Microsoft Teams, SharePoint, Adobe, Figma, Snowflake, and WhatsApp 1

1

2

2

. GoDaddy reportedly eliminated 15,000 hours of manual work annually using Quick 3

3

.

Kiro mobile app and infrastructure updates support continuous agentic AI operations

Kiro, AWS's specification-driven AI coding assistant, now offers a closed preview iOS mobile app that lets developers launch and manage remote sessions from their phones 1

1

. The native app features three interaction modes—chat, spec for specification workflows, and autonomy for delegating tasks—and renders code diffs as cards designed for small screens 1

1

. Deepak Singh, AWS VP leading the Kiro team, noted that faster AI code generation creates more work for humans to review, test, and maintain, describing these as "good problems to have, but real problems" 2

2

. AWS also introduced AWS Context, a service that maps company data into knowledge graphs for agentic search, publishing metadata into Amazon S3 tables in Apache Iceberg format with identity-aware queries 1

1

. Bedrock AgentCore, the platform for custom agents, now includes managed knowledge bases, web search, and the ability for agents to access paid content like financial market feeds 1

1

.

Governance emerges as critical factor in autonomous agent deployment

Sivasubramanian framed security and governance as a false choice between "walled gardens" that limit agent capabilities and "wild gardens" that lack enterprise controls 3

3

. Quick addresses this by ensuring every action carries its own governance, tracking who acted, what data they touched, where it went, and whether policy allowed it 3

3

. For IT professionals, this shift means architectural decisions over the next 12 to 18 months will determine whether autonomous agents become force multipliers or sources of chaos, with policy as code, identity boundaries, and least-privilege design becoming as critical as model selection 3

3

. The simultaneous release of autonomous capabilities and extensive guardrails suggests AWS recognizes the tension between selling effortless autonomy and the reality that enterprises need tools to monitor, second-guess, and potentially undo agent actions 4

4

. Wood claimed that while frontier token costs continue rising, the cost normalized for a particular level of intelligence decreases year over year, though services like Quick use subscription pricing and DevOps Agent charges per-second usage fees 1

1

.