Inside AWS’s plan to use agentic AI to reset legacy IT

Agentic AI was the headline concept across keynotes, briefings, and product launches at AWS re:Invent.

Systems that reason, act, learn, and execute across workflows are now being positioned as the next phase of enterprise automation. Yet inside most organizations, the same question still lands in every boardroom. What do these announcements actually mean for the systems that keep businesses running every day?

AWS used re:Invent 2025 to provide a direct answer. The company did not build its message around futuristic demos. Instead, it focused on something enterprise leaders know all too well. Technical debt. Years of legacy systems, Windows estates, mainframes, VMware environments, and brittle integrations that slow every transformation effort. AWS presented agentic AI as the toolset that finally makes large-scale modernization possible at speed.

Across AWS Transform, Nova, Nova Forge, Bedrock AgentCore, Frontier Agents, AI Factories, and its partner ecosystem, AWS revealed a single connected narrative. Reset the past, automate the present, and use agentic systems to reshape how enterprise software is built and operated.

Tech debt – a barrier that AI cannot ignore

AWS placed technical debt at the centre of its story. According to figures the company shared, a typical organization spends about 30 percent of its teams' time on manual modernization work. This effort is required to keep systems alive, yet it drains momentum from projects that create new business value.

AWS Transform has already analyzed more than 1.1 billion lines of code and removed over 810,000 hours of manual effort across customer projects. Customers using earlier transformation capabilities have modernized systems four times faster than with traditional approaches.

During re:Invent, AWS framed tech debt as the hidden weight holding back enterprise AI. Large language models, automation stacks, and intelligent assistants struggle when connected to ageing systems built for a different era. This framing moves the AI conversation away from novelty and grounds it in infrastructure reality.

Even the symbolism reflected that shift. AWS staged a demolition of decommissioned server equipment at the event to underline the message. Legacy systems no longer need to remain frozen in place. Agentic systems are now positioned as the tools that can dismantle, reassemble, and modernize at scale.

AWS Transform shows how agentic AI is applied to real systems

The centrepiece of the week was the expansion of AWS Transform. AWS describes it as the first agentic AI service explicitly built for large-scale applications and code modernization. Transform now spans Windows .NET applications, SQL Server, user interfaces, operating systems, mainframes, VMware, APIs, legacy runtimes, and even organization-specific languages.

AWS stated that Transform can accelerate complete Windows modernization by up to five times and reduce maintenance and licensing costs by up to 70 percent. The company backed those figures with live customer examples.

Air Canada modernized thousands of Lambda functions in just a few days, reducing the expected time and cost by 80 percent. QAD reduced customer upgrade cycles from two weeks to three days, delivering productivity gains of up to 70%. Thomson Reuters now migrates 1.5 million lines of code per month and has reduced technical debt by 50%.

In an interview, Madhu Parthasarathy, general manager of AgentCore at AWS, reinforced how far this shift has already moved.

"Several thousands of hours reduced into weeks to migrate legacy systems," he said.

This is where agentic AI separates itself from standard automation. Transform captures outcomes, improves with each project, and scales across entire estates. Modernization shifts from a multi-year initiative into continuous operational work.

"We are seeing orders of magnitude difference in productivity increases," he said.

This reframes how CIOs approach transformation. Instead of pausing innovation while legacy systems are rebuilt, agentic systems compress that work into the flow of daily operations.

For enterprise leaders, this shift changes planning cycles, staffing models, and financial forecasting. Modernization stops being a front-page initiative and becomes part of the business's background rhythm.

Mainframe and VMware estates move closer to automation

AWS extended AWS Transform into mainframe and VMware environments. On the mainframe side, new agents now handle activity analysis, business rule extraction, documentation generation, and automatic test planning. These are the stages that have historically stalled migrations and driven up project risk.

The powerful new agentic AI capabilities also make it easier to automate VMware migrations to AWS, including support for security and networking stacks from Cisco ACI, FortiGate, and Palo Alto Networks. A new on-premises discovery tool strengthens inventory and security review across estates.

For enterprises running critical workloads across these platforms, the implication is clear. Dependency mapping, documentation, and migration planning are no longer strictly manual workstreams. Agentic systems now handle a large share of that load.

AWS also introduced three Frontier Agents at re:Invent. There’s Kiro, which is an autonomous development agent. There’s also an AWS Security Agent to support application design and testing, as well as an AWS DevOps Agent focused on reliability, prevention, and response.

These agents operate for extended periods with context awareness and minimal human input. They are built to reason, execute, and refine continuously across development and operations.

For enterprises, this points to a structural shift in how software teams function. Engineers direct intent and supervise outcomes. Agentic systems handle the repetitive, time-consuming layers of execution.

The fundamental shift from prototypes to production

The most critical enterprise signal from the event came from AWS's view of the adoption curve itself.

"We built Agent Core to help customers fill the chasm from prototyping an agent to actually deploying them in production," Madhu told me. He went further by tying that shift to the coming year.

"I feel 2026 is going to be a year where we will see a lot more agents being deployed in production and operating at scale. That shift is starting to happen right now."

This explains why Bedrock AgentCore became such a focal point at re: Invent. The transition from experimentation to operational reliance is the moment enterprises care about most.

It’s also important to note that technology is no longer the leading blocker. But Trust is. "One of the key impediments for a lot of enterprise customers is the reluctance to see how agents behave in production," Madhu said.

This is where the new policy layer inside Bedrock AgentCore becomes central to adoption.

"The policy announcement becomes so much more valuable because customers finally have a way to control what agents can do outside the purview of the agents," he added.

He described the policy system in a way that resonates strongly with enterprise risk teams.

"That provides determinism in a non-deterministic system."

Policy enforcement, evaluations, and episodic memory now give enterprises visibility into how agents act, learn, and improve over time. Customers across healthcare, finance, manufacturing, and media are already using these tools to move from controlled pilots into production deployments.

Multi-agent orchestration becomes a new design pattern

Rather than relying on fixed workflows designed entirely by human designers, enterprises are beginning to hand over orchestration decisions to reasoning models.

"Rather than hard-code orchestration workflows, we can leverage models to reason about orchestration steps," Madhu continued.

This represents a bigger structural change in enterprise software design. Static automation gives way to adaptive orchestration driven by contextual reasoning across multiple agents.

AWS positioned the Nova model family as the intelligence layer powering this ecosystem. Nova 2 models now cover reasoning, speech, and multimodal workloads with built-in web grounding and code execution.

Nova Forge then extends that capability into custom training. Enterprises can blend proprietary data across multiple stages of the training pipeline, build domain-specific variants, run reinforcement learning in simulated environments, and deploy those models inside Bedrock with full security controls.

Madhu summarized the importance of this shift simply: "Your data is compelling, and it is everything." This signals a move away from universal intelligence toward organisation-owned model layers shaped by internal knowledge.

Nova Act was revealed as a service that automates browser-based workflows. Built on Nova 2 Lite and trained through reinforcement learning across hundreds of web environments, it enables enterprises to automate CRM updates, testing, claims processing, and other UI-driven tasks through agent fleets.

For enterprises, this opens a new layer of automation across systems that were previously resistant to API based integration. Early adopters, including Hertz and 1Password, have already used it to compress quality assurance and deployment cycles from weeks into hours.

Looking at the bigger picture, AWS is going toe-to-toe with Google’s Gemini as an "all-in-one" offering for text, speech, images, and video.

Partners accelerate the agentic ecosystem

Another consistent signal at re: Invent was the expanding role of AWS partners. Research shared at the event showed that partners now generate more than $7 in services revenue for every $1 spent on AWS technology. More than 80 percent of partners now deliver AI solutions as part of their transformation work.

AWS introduced new agentic AI partner categories covering applications, development tools, and consulting services. Marketplace updates added AI-powered discovery, automated private offers, and multi-vendor solution bundles.

For enterprises, this means procurement, deployment, and long-term optimization of agentic systems will increasingly flow through partner-led models.

Taken together, the announcements at re: Invent 2025 show AWS consolidating its AI strategy around agentic execution. AWS Transform handles legacy modernization. Nova supplies reasoning. Nova Forge enables custom intelligence. Bedrock AgentCore provides memory, policy, and evaluation. Frontier Agents reshape engineering and operations. AI Factories extend compute into regulated environments. Partners package these capabilities into services.

For enterprises, this signals a shift where AI is no longer confined to analytics or experimentation. It is entering the core layers of IT, development, security, and infrastructure itself.

During the California Gold Rush, the most profitable businesses were those that sold supplies and services to miners, not those that mined gold. One hundred eighty years later, we are experiencing an Agentic AI gold rush, and Amazon and its partners want to be the ones selling virtual picks and shovels to enterprises.

Unlock Cybernews content on YouTube