Amazon AWS: Three New AI Agents for Everyday Development

At its annual re:Invent trade show, AWS made it unmistakably clear where the journey is heading: Agentic AI systems are intended to not only handle simple tasks in the future but to work autonomously for hours or even days. With the so-called Frontier Agents, Amazon announced a new generation of AI agents that are supposed to maintain persistent contexts and manage complex workflows without constant human guidance.

The focus is on three specialized agents designed to transform the software development cycle. The Kiro Autonomous Agent acts as a virtual developer, processing backlogs, classifying and prioritizing bugs, and independently solving tasks across multiple code repositories. It continuously learns from feedback and pull requests. The AWS Security Agent takes on the role of a virtual security consultant: it reviews design documents and pull requests for vulnerabilities, adheres to organization-specific requirements, and turns time-consuming penetration tests into an on-demand function. Completing the trio is the AWS DevOps Agent, which, as part of the operational team, diagnoses incidents, correlates telemetry and deployment data, and proactively suggests improvements.

Bedrock AgentCore is Extended

The central platform for operating these agents, Amazon Bedrock AgentCore, received three significant extensions. With AgentCore Policy, policies formulated in natural language, known as guardrails, can now be enforced, blocking unauthorized agent actions in real-time. AgentCore Evaluations offers 13 pre-built metrics for quality assurance, such as for correctness and contextual relevance. AgentCore Memory, an episodic memory, allows agents to learn from past interactions and adjust their decision-making.

The portfolio is complemented by Amazon Nova Act, a service for automating browser UI workflows, which AWS claims achieves 90 percent reliability. The open-source AI agent SDK Strands Agents, which AWS introduced just this May and initially focused on Python, has also been extended with TypeScript support and now runs on edge devices for automotive and robotics applications.

Warning of Exploding Costs

Despite all technical sophistication, the economic side of agentization has a significant drawback. Jeff Boudier, Chief Product & Growth Officer at Hugging Face, explains in an interview with iX that the use of AI agents must fundamentally be questioned. The transition from classic LLM applications to agentic systems leads to drastically higher computational effort due to their iterative working method.

"Instead of a user request costing about one cent, with agentic systems we quickly reach three, five, or even more dollars," Boudier explains. "This shift will trigger an enormous amount of engineering work in the coming year to reduce these costs again."

Therefore, companies must first ask themselves whether they actually need an agent for a specific use case, what value a correct answer has, and what the risk is if an agent makes wrong decisions. Whether AWS's promised reliability and scalability of their agents justify the additional costs will also have to be seen in practice.

(dahe)