Why Enterprise Agents Perish at the Prototype Stage; AWS's Chu Ruisong: Agent Engineering Is the Decisive Factor

"The inflection point for the Agentic AI explosion has arrived," declared Chu Ruisong, Amazon Global Vice President and AWS Asia Pacific President. The catalyst for this shift is twofold: the sustained evolution of foundation models and, more critically, the accelerating maturity of the Agentic engineering framework.

Ganapathy "G2" Krishnamoorthy, AWS Vice President of Global Database Services, further observed that Agentic AI's defining transformation is its evolution from a single-turn Q&A utility into a persistent digital workforce—one capable of sustained context comprehension, tool invocation, task execution, and continuous learning. Competitive advantage will accordingly pivot from "AI adoption" to "the ability to convert each task into intelligent capital for the next."

Future AI Agents, he noted, will span several critical domains: agents redefining workforce workflows, agents strengthening security and compliance posture, agents capable of software comprehension and delivery, and platform-level capabilities enabling enterprises to build their own agentic systems.

Over the past two-plus years, Chu Ruisong argued, the foundational underpinnings of Agentic AI have undergone material shifts. Foundation model capabilities—spanning reasoning, code generation, and multimodal comprehension—have crossed successive intelligence thresholds amid intensifying competition. Simultaneously, the Agentic engineering framework built atop these models has accelerated toward maturity.

"The two are forging a self-reinforcing flywheel: advancing model capabilities equip agents with superior reasoning, code generation, and multimodal faculties; mature Agentic engineering practices, in turn, channel real-world business feedback into the model iteration loop, steering subsequent breakthroughs."

Chu defined the Agentic engineering framework as the systematic discipline of translating model capabilities into production-ready agents that consistently yield business outcomes. It is not a discrete technical artifact but a holistic engineering architecture encompassing models, context, tools, workflows, evaluation, and governance.

The Agentic engineering framework can be conceptualized as a three-tier architecture.

On enterprise Agentic transformation, Chu advised that organizations must first pivot from a tool-centric to an outcome-centric paradigm. Historically, AI initiatives fixated on model selection, tool procurement, and system architecture. In the Agentic AI era, however, the defining question is: what business outcome is AI intended to produce? Only once the business objective is unambiguously defined can model selection, data readiness, workflow design, and platform engineering proceed with strategic clarity.

For enterprises initiating their first agent project, Chu advised that the initial use case need not be complex, but must carry genuine business value. An ideal candidate scenario features unambiguous start and end points, a defined business objective, cross-system and cross-tool decision-making requirements, and measurable success metrics. Crucially, it must also incorporate a safe failure mode—agent errors should not trigger irreversible consequences. Enterprises must further delineate the agent's boundaries: its degree of autonomy, jurisdictional scope, and the human-in-the-loop review framework.

A practical methodology, Chu proposed, is to define an agent as one would draft a job description for a new hire: articulate its responsibilities, delivery benchmarks, authority limits, and error-handling protocols—then tie each directly to business KPIs. This enables the enterprise to measure business value from the very first agent workflow and systematically scale across additional Agentic AI workflows.

Second, data is emerging as one of the enterprise's most formidable moats. In the Agentic AI era, data transcends its legacy role as a static repository; it becomes a strategic asset that perpetually fuels agent-generated value. Decades of accumulated industry knowledge, customer data, operational workflows, and business expertise will constitute the least replicable dimension of an enterprise's agentic competitive advantage.

Simultaneously, the Agentic platform will constitute the fault line between proof-of-concept and production-scale deployment. For a handful of agent pilots, localized tools and point solutions suffice. But when hundreds or thousands of agents must operate in concert with human employees—and with one another—a unified platform becomes indispensable for development, deployment, management, evaluation, and governance. Platform capability is the gatekeeper between experimentation and industrial-scale production.

Furthermore, Chu stressed that trust and governance function as accelerants, not impediments, to enterprise progress. New roles, processes, accountabilities, and management frameworks must be established to orchestrate the human-agent collaborative dynamic.

"For many enterprise employees, the current mode of working is already increasingly untenable for effective collaboration," Krishnamoorthy observed. Employees are indifferent to which tool a task lives in; their genuine concern is whether an idea can be pushed forward, a customer issue resolved, or a change executed. In practice, however, they are forced to shuttle between disparate systems, performing the function of "information broker, decision maker, and workflow driver."

A more fundamental problem: many AI tools, though capable of integrating with enterprise systems, answer a single question and discard the context. They operate in siloed sessions, processing isolated requests from singular data sources—unaware of the user's collaborators, ongoing projects, or the broader business workflow they inhabit. They cannot weave fragmented processes into a coherent intelligent system.

Krishnamoorthy contends that as long as teams rely on humans to bridge information gaps, and users remain forced to serve as manual workflow operators, the efficiency gains AI can unlock will remain severely constrained. Leading organizations accelerate not merely through faster execution, but by engineering each task to build intelligent capital for the next.

A genuinely transformative agent must operate across enterprise business domains, data ecosystems, and toolchains—learning continuously with each query and interaction. Consequently, the answer a user receives on day 100 should be demonstrably more accurate and contextually relevant than the one delivered on day 1. This elevates the agent beyond a conversational interface: it must embed persistent memory, permission governance, contextual reasoning, and cross-system execution. It must track the acting entity, the data accessed, the provenance of that data, and whether the prevailing policy sanctions the operation.

Enterprises have long grappled with a fundamental tension: security and speed appear inherently adversarial. Rapid development, deployment, and customer responsiveness sit on one side; security review, vulnerability scanning, compliance verification, and risk control on the other. Historically, security has been cast as the drag on velocity.

In the Agentic AI era, however, this trade-off is no longer tenable. Krishnamoorthy argues that enterprise security is transitioning from a paradigm of "storage, query, dashboard, and alert" to one of "context, reasoning, and access control." Security signals stripped of context are mere noise. Only when informed by an organization's specific environment, architecture, entitlements, and business logic can a security system make risk-priority judgments with authority.

Beyond workflows and security, software development constitutes another critical vector for Agentic AI deployment. Krishnamoorthy noted that virtually every agent, workload, and business touchpoint will ultimately be delivered as software to customers. Software delivery, therefore, is the pivotal mechanism through which enterprise innovation velocity is converted into tangible business value.

Code generation has become increasingly cheap, and tools capable of rapid code synthesis are abundant. For enterprise-grade projects, however, the bottleneck is not code authoring—it is comprehending the enterprise environment, development conventions, system architecture, and legacy constraints. A code generator lacking enterprise context risks accruing technical debt rather than enhancing engineering productivity.

AWS contends that enterprises require not point-solution code generation, but end-to-end software delivery capability spanning requirements comprehension, system design, implementation, testing, verification, and release. An agent must understand project structure, architectural constraints, and testing criteria, and must complete pre-release validation to ensure code is not merely "generable" but "correctly shipable."

As agents accelerate code production, the software pipeline must evolve in lockstep. If an agent can produce code at ten times the velocity while the release pipeline remains anchored to legacy throughput, the bottleneck merely migrates—from "code authoring" to "testing, review, and deployment."

Consequently, release management itself must become agentic. Post code generation, the system should autonomously orchestrate the downstream pipeline: reviewing changes, provisioning environments, executing tests, detecting failures, remediating issues, and completing validation before any problem impacts the customer.

Future software delivery will transition from "humans pushing the pipeline" to "agents orchestrating the pipeline." Multiple agents can collaboratively sustain a continuously flowing development, testing, and release cycle, enabling enterprises to accelerate delivery without the compounding weight of technical debt.

AWS further observed that many enterprise agents today remain trapped at the prototype stage, unable to reach production. The impediment is not insufficient model capability but the absence of production-grade agent infrastructure: authentication, persistent memory, long-running task execution, security, governance, tool connectivity, policy enforcement, and a managed runtime environment.

From AWS's perspective, the enterprise's focus should extend beyond the agent itself to the full platform that sustains production-grade agent operations. The model functions as the "brain"; the harness or platform capabilities constitute the "body" and "security apparatus." Absent the latter, even the most capable model will struggle to reliably execute enterprise-grade tasks.