NYXN has developed Forge, a methodology designed to help companies adopt AI agents while maintaining control over security, costs, and outcomes. The approach combines business context, governance, observability, and human oversight, and is being applied to modernize a banking core in an estimated 11 months with 12 people and 47 AI agents.

 

The shift from generative AI to agentic AI is changing how companies approach software development and business operations. While GenAI has largely been used to produce content or accelerate individual tasks such as code generation, AI agents can connect to enterprise systems, use tools and execute workflows across multiple stages. That distinction is becoming increasingly important as companies look to capture productivity gains without creating new risks around security, autonomy, cost and accountability.

NYXN, a company specializing in enterprise adoption of agentic AI in Latin America, developed Forge as a methodology for integrating agents into technology and business processes. The company is applying the approach to the modernization of a banking core, a project estimated to reduce the development timeline from three years to 11 months with a team of 12 people and 47 AI agents.

The project illustrates the complexity of applying AI to legacy enterprise environments. According to David Casillas, Chief Technology Officer, NYXN, the banking core involved in the project contains more than 2,000 stored procedures, with approximately 14,000 lines of code each, developed over 15 to 20 years. Such systems do not contain only code. They also incorporate business rules, historical decisions and internal annotations that may reflect regulations or operational requirements that were implemented years ago. Much of that knowledge can be difficult to reconstruct because the engineers who originally developed the system may no longer work for the organization and documentation may have evolved over time.

“The complexity in modernizing a core is high or very high, because you have to take that code, understand it, see what documentation was created at the time and reconstruct rules that in many cases are only in the heads of the engineers who originally programmed them,” says Casillas.

AI agents provide a different way of addressing that complexity. Instead of requiring engineers to manually read and interpret thousands of lines of legacy code, agents can process large volumes of information, identify relationships, and support the analysis and transformation of the existing system.

From Code Generation to the Full Engineering Cycle

NYXN’s experience suggests that the largest productivity gains from agentic AI do not necessarily come from writing code faster. A conventional software development process includes multiple stages: product and architecture design, code generation, testing, and quality assurance, deployment, security configuration, and operation. GenAI can accelerate code creation, but the rest of the process can continue to generate bottlenecks.

“Most agents today are being used to create code, but the development process of any technology application has many phases,” says Casillas. “What happened with GenAI was that we only attacked the code-creation part.”

According to Casillas, code generation alone can produce a relatively limited impact because teams still need to design the application before coding and test, secure, deploy and operate it afterward. NYXN estimates that the maximum time savings from code generation alone can be around 12% of a project’s total timeline.

Forge instead seeks to “agentize” the complete engineering cycle, from design and code generation through QA, deployment, security, and operations. This broader approach is where NYXN says it has observed its largest gains. Over two years, the company compared projects developed with AI agents against projects using traditional agile methodologies and identified execution-time improvements of between 50% and 70%.

Casillas says testing and QA have emerged as particularly important areas for agent deployment. Agents can test applications, identify errors, return those findings to the development process, and support subsequent corrections, reducing the back-and-forth traditionally required between development and testing teams. The objective is not simply to make individual engineers faster but to remove bottlenecks throughout the sequence of activities required to move software from design to production.

Context Becomes a Critical Layer

Forge is also based on the premise that AI agents cannot operate effectively without sufficient business and technical context. Rather than treating an agent as a system that receives a simple prompt and produces an output, NYXN structures the information required to execute a task, including business objectives, rules, technologies, architecture and quality requirements.

That context becomes particularly important in legacy environments, where critical business logic may not exist in formal documentation and can instead be embedded in code accumulated over years. For NYXN, this makes context a functional component of agentic engineering rather than supplementary information.

The governance layer is designed to determine how agents can be used, what permissions they receive and which actions require additional controls. Forge uses the AI Management System framework defined by ISO/IEC 42001 as a reference for this governance approach.

Casillas compares the model to the separation of powers in government: organizations establish rules for how AI can be used, Forge provides the methodology for executing those rules and observability mechanisms monitor what happens in practice. That monitoring includes the actions agents perform, the models they use, the number of execution cycles, token consumption, and the cost associated with individual tasks.

Security is another key consideration. Because agents can interact with enterprise systems, they may require credentials, permissions, and access to applications or databases. Improperly configured access or malicious manipulation of an agent’s instructions could therefore create risks beyond those associated with conventional software tools.

Human Oversight Defines the Boundaries

Forge’s “Human on the Loop” principle addresses another question raised by autonomous systems: who decides what an agent is allowed to do? Rather than establishing a universal list of tasks that agents can or cannot execute autonomously, NYXN places that responsibility with the professional accountable for the work.

“For us, agents are a tool that extends what a professional knows and does,” says Casillas. “The agent is going to do what the human tells it to do.”

Under the model, organizations establish rules at the beginning of a project defining which actions agents can execute independently and which require authorization. An additional agent can monitor those activities and verify that permissions and approval requirements are being followed. The human therefore remains responsible for understanding the task, defining the instructions and determining the level of autonomy that is appropriate.

The approach can extend beyond banking. NYXN sees agentic engineering and business-process automation as applicable across industries, with the longer-term objective of helping organizations develop what it calls an “AI mindset.” Rather than deploying isolated AI projects, companies could progressively operate with hundreds of agents distributed across technology and business processes.

From Agentic Projects to Agentic Companies

For NYXN, Forge’s long-term role is to provide a repeatable framework for this transition. The first stage is to enable companies to agentize technology projects in a structured way. The second is to extend the same principles to business processes, with the objective of making those processes scalable, secure, measurable and economically viable.

The broader shift is therefore from using AI as a productivity tool toward restructuring how organizations operate around autonomous systems. “The intention of Forge is to turn companies into agentic companies, or what we call an AI mindset, where through the use of AI these companies become more profitable, more scalable and more efficient over time,” says Casillas.

For enterprises evaluating agentic AI, the implication is that the central question is moving beyond what individual models can generate. The more consequential question is how organizations can structure context, permissions, workflows, observability, and human accountability so that agents can operate across the enterprise without losing control of the systems they are designed to accelerate.