The Prompt is a Relic of Wave One AI

The Fallacy of the Perfect Prompt

For the last three years, the global obsession has been prompt engineering. We treated the Large Language Model (LLM) as a temperamental oracle, believing that the right combination of adjectives and constraints could unlock peak performance. This approach, characterized by Tirias Research as Wave 1, viewed AI primarily as a conversational assistant responding to individual, isolated prompts. But this paradigm is collapsing. The industry is realizing that while a prompt can generate a clever response, it cannot execute a business process. The bottleneck has shifted from how we talk to the model to how the model interacts with the world.

We are now entering Wave 2, where the unit of value is no longer the response, but the outcome. In this new phase, autonomous agents reason, invoke tools, maintain long-term context, and execute multi-step tasks with minimal human intervention. When an AI system is tasked with reconciling a quarterly budget across four different regional offices in Southeast Asia, a prompt is insufficient. What is required is a system that can navigate data silos, handle errors autonomously, and verify its own work. The prompt is not the engine; it is merely the ignition switch.

This shift fundamentally redefines the technical challenge. Running a model is largely a compute problem—a matter of H100 clusters and energy efficiency. However, running an agent is a systems problem. It requires a platform that combines compute, models, data, tools, and operational services into a unified environment. If the model is the brain, orchestration is the nervous system. Without it, the brain is an isolated organ, capable of thinking but unable to move or act upon its environment.

The Emergence of the Control Plane

As enterprises attempt to move beyond proof-of-concept, a new product category has emerged: the agent gateway. These gateways serve as the control plane for enterprise AI, sitting between the autonomous agent and the tools or models it accesses. The goal is to move away from the chaotic 'wild west' of agentic autonomy toward a governed environment. Nutanix has already integrated this into its Enterprise AI 2.7, providing a centralized point to manage traffic from agents to LLMs and business tools.

The necessity of this layer becomes clear when considering the risks of unbridled autonomy. Arcade has recently made its agent authorization and tool-execution runtime available through Microsoft Azure and AWS marketplaces, allowing companies to deploy these controls within their own cloud. Why is this critical? Because an agent that can call an API to move funds or delete records cannot be managed by a prompt. It must be managed by a gateway that enforces permissions, tracks costs, and ensures ownership.

This sentiment highlights the gap between generic model capabilities and business utility. To bridge this, we are seeing the rise of specialized knowledge layers. Adverity Atlas, for instance, provides a governed understanding of marketing data, sitting on top of existing data warehouses. This is not about prompting the AI to be a better marketer; it is about providing the AI with a structural understanding of the business context. It is the difference between giving a contractor a manual (the prompt) and giving them the actual blueprints of the building (the knowledge layer).

The danger of ignoring this systems-level shift is quantifiable. Gartner warns that more than 40% of agentic AI projects could be canceled by 2027. Crucially, these failures will not be caused by poor model quality. Instead, they will stem from management issues. Many organizations are currently deploying what are essentially chatbots with ambitions—systems that can talk about doing a task but lack the orchestration layer to actually execute it safely and reliably at scale.

The Security Paradox of Autonomy

Autonomy introduces a terrifying new attack vector: the agentic threat actor. We have already seen the first documented end-to-end agentic ransomware operation, dubbed JadePuffer. This AI agent did not simply follow a human's script; it autonomously exploited a vulnerable Langflow server, harvested credentials, moved laterally through the network, and encrypted over 1,300 database records. JadePuffer adapted its actions on the fly, executing more than 600 coordinated payloads to achieve its goal of extortion.

This incident proves that traditional security is obsolete in the face of agentic AI. Static credentials and standing privileges—the bedrock of current IT security—are insufficient when an agent can authorize, limit, and revoke permissions multiple times within a single workflow. If an agent has a permanent key to the kingdom, a single exploit allows a machine-speed intrusion that no human security team can counter in real-time.

• Dynamic Lifecycle Identity: Moving from static keys to temporary, task-specific certificates.
• Agent-to-Agent Governance: Restricting which autonomous agents are permitted to communicate with one another.
• Privileged Access Management: Implementing just-in-time permissions that expire the moment a task is completed.
• Non-Human Identity (NHI) Frameworks: Treating agents as distinct entities with their own audit trails and accountability logs.

The solution is a dynamic, lifecycle approach to agentic identity. Organizations must establish a reliable identity for agents, treating them as non-human identities similar to service accounts but with far more granular control. The goal is to create a 'certificate' of identity that is recognized and governed across different cloud environments. Without this, the orchestration layer becomes a liability rather than an asset.

Ultimately, the transition from Wave 1 to Wave 2 is a transition from art to engineering. Prompting was an art form—a search for the magic words. Orchestration is engineering—the construction of gateways, knowledge layers, and identity frameworks. Those who continue to invest primarily in the 'prompt' are building on sand. The real value is being captured in the invisible layer: the systems that tell the AI not just what to think, but how to act, where to stop, and who to answer to.

As we look toward 2027, the divide will widen between companies that have 'AI tools' and companies that have 'AI systems.' The former will struggle with the 40% failure rate predicted by Gartner, trapped in a cycle of endless prompting and fragile pilots. The latter will have built a robust control plane, turning their LLMs into a workforce of autonomous agents that are governed, secure, and deeply integrated into the business logic of the organization.