AWS customers now face a dual imperative: extract measurable efficiency from existing cloud spend while simultaneously rebuilding development workflows around autonomous AI agents. The release of the AWS State of Cost Efficiency Report, based on analysis of more than 71,000 opted-in accounts, supplies the first large-scale benchmark for how organizations actually achieve optimization. At the same time, a cluster of new tools for building, evaluating, and operating agents signals that frontier teams are already treating AI as infrastructure rather than an accessory.
This convergence matters because cost discipline and AI adoption are no longer sequential initiatives. Teams that adopt agentic patterns without visibility into their spend risk amplifying waste; those focused solely on rightsizing may miss the 4.5x to 10x productivity gains reported by organizations that have restructured work around AI-native practices. The data and tooling released in recent weeks make both dimensions visible and actionable.
Cost Efficiency Benchmarks Reveal a Persistent Optimization Gap
The Cost Efficiency metric, introduced at re:Invent 2025 and now tracked daily in Cost Optimization Hub, combines rate optimization through Savings Plans and Reserved Instances with workload actions such as rightsizing and idle-resource removal. Across the analyzed customer base, the median score stood at 83 while the mean reached only 79, a divergence driven by a long tail of under-optimized accounts.
Larger customers that combine Savings Plans with active rightsizing run roughly 60 percent of their EC2 instances on newer hardware and improve their median score four times faster than peers relying on Savings Plans alone. High Savings Plans coverage, however, can obscure remaining opportunity: organizations at 95–100 percent coverage see non-Savings-Plan optimization potential drop 65–80 percent relative to those below 25 percent coverage. Enabling EC2 memory metrics correlates with 8–30 percentage-point gains in realized savings, yet only 17.7 percent of eligible customers have activated the feature.
These figures indicate that the largest remaining efficiencies lie in operational visibility rather than rate negotiation. The report underscores that Savings Plans coverage alone is insufficient; sustained improvement requires continuous instrumentation and customized recommendation tuning within Compute Optimizer.
Frontier Teams Restructure Development Around AI Agents
Separate research into AI-native development at Amazon shows that productivity leaps occur only when teams redesign workflows rather than layer agents onto existing processes. One Bedrock team of six engineers completed a project originally scoped for 30 developers over 12–18 months in just 76 days. Over five months the same group shipped more production code than the preceding decade.
The decisive variable is not code-generation speed but agent access to institutional knowledge and the willingness to retire coordination overhead. Pathfinder initiatives, structured sprints, and controlled in-situ experiments across hundreds of Amazon teams converged on the same pattern: successful groups treat AI adoption as an engineering investment that retires technical debt and reduces non-coding tasks. The productivity gap between these frontier teams and others is widening rapidly, with commits rising while production feature velocity remains flat for teams that have not restructured their operating model.
Systematic Evaluation Closes the Reliability Gap for Agents
As agent deployments proliferate, output-level testing proves inadequate for systems that autonomously select tools and maintain state across multiple steps. An agent may produce a coherent final answer while hallucinating from empty tool results or skipping required verification steps. Agent-EvalKit, an open-source toolkit released under Apache 2.0, addresses this by tracing full execution paths—tool calls, intermediate state, and faithfulness to source data—within the developer environment.
The framework generates targeted test cases from natural-language goals, instruments observability, and surfaces improvement recommendations tied to specific code locations. When applied to travel-research agents built with Strands Agents and Amazon Bedrock, it surfaces failure modes invisible to conventional unit tests. Widespread adoption could shift agent development from post-deployment firefighting to disciplined engineering practice.
Automation Reduces Friction in Database and Messaging Migrations
Operational efficiency also extends to legacy modernization. Amazon MQ version 4 now supports the RabbitMQ JMS Client and Topic Exchange plugin, enabling existing JMS 1.1–3.1 applications to migrate from on-premises ActiveMQ or self-managed brokers with only connection-factory changes. The approach eliminates weeks of rewrite work while preserving publish/subscribe semantics through server-side routing.
Parallel automation for Aurora PostgreSQL upgrades uses AWS Systems Manager and EC2 to orchestrate readiness checks, Copy-on-Write clone creation for rollback safety, and version promotion across tagged fleets. Organizations report up to 80 percent reduction in manual effort. These patterns demonstrate that mature automation layers can convert high-risk, infrequent operations into repeatable, low-touch processes.
Real-Time Data Pipelines and Semi-Structured Storage Converge
The ability to keep analytical copies current without batch exports is advancing through change-data-capture pipelines that land directly in Apache Iceberg tables on Amazon S3. Using Debezium on MSK Connect, events from Aurora PostgreSQL are transformed and delivered via Firehose so that row-level inserts, updates, and deletes appear as native Iceberg operations. The destination tables remain query-ready for Athena, Redshift, and SageMaker without downstream reconstruction of current state.
Complementing this capability, Apache Iceberg V3 introduces the VARIANT data type. Rather than storing evolving JSON payloads as opaque strings, VARIANT shreds documents into metadata, a binary fallback, and typed columns. Queries that target specific nested fields scan only the relevant sub-columns, reducing both CPU time and storage footprint. Together these technologies narrow the traditional divide between operational and analytical workloads.
The developments described above share a common thread: AWS is supplying both the measurement frameworks and the execution tooling that allow organizations to treat cost, development velocity, and data freshness as integrated engineering concerns. Teams that align their agent adoption strategies with the visibility provided by Cost Optimization Hub, while adopting systematic evaluation and automated migration patterns, position themselves to capture compounding gains rather than isolated improvements. The next phase will likely be measured by how quickly these practices move from early adopters to the broader installed base.
