The
era of agentic development has raised new questions about what is best for creating, maintaining, and evolving the software of today and tomorrow.

But before debating which option is best – whether GeneXus, Globant Coda, Claude Code, Codex, or any other tool currently available – it is worth understanding what problem each one solves, what their scope is, and where their limits appear. That perspective allows us to have a more honest conversation, less driven by hype and more connected to the real needs of enterprise development.

Code First or Knowledge First

If you are building a small application, a prototype, an automation, a script, a proof of concept, or a product where you yourself will control the repository, Claude Code will probably give you a more immediate experience than GeneXus.

This is because GeneXus does not compete solely on the field of “who generates a nice prototype fast” or “who can create more lines of code more quickly.”

GeneXus’s strength appears when the problem stops being “generate code fast” and becomes building, maintaining, and evolving complete enterprise systems: with data models, business rules, APIs, interfaces, security, integrations, traceability, deterministic generation, governance, and a long lifecycle.

Claude Code works very well on code. You can use specs, documents, or some spec-driven development approach to try to preserve part of the knowledge of what you are building, all using Generative AI.

GeneXus works on a knowledge base first and generates afterward. It starts from the structured knowledge of the system (which we call the Knowledge Base), which contains data, rules, relationships, etc., and from there generates code – but it does so with a different type of generator, a form of deterministic (symbolic) AI to model, infer, and generate software consistently.

It is an analogous strategy in its objective – accelerating software creation – but very different in its nature.

The
New Bottleneck: Verifying What Was Generated

Code generation through generative AI does not seem to have limitations at first glance, because whatever we ask it, it will try to do it. It does not matter the language, programming language, or the request we make. But coding agents also have a limitation that becomes evident when we look at where the bottleneck has shifted.

Before, the bottleneck was creating code. Now, increasingly, the bottleneck becomes verifying what was generated: checking whether the code does what it was supposed to do, whether it respects the architecture, whether it does not break something existing, whether it does not introduce vulnerabilities, whether it is maintainable, whether it fits with the rest of the system, and whether it will still be understandable six months from now.

In small applications, that overhead is totally tolerable. It may not even matter too much. If something breaks, the same developer who generated the code probably understands it, fixes it, and moves on.

But in large applications, with several teams, many business rules, integrations, compliance, security, critical data, and years of evolution ahead, that control cost starts to weigh.
The
generation speed stops being the only relevant indicator.
The
questions become different: How much effort do I need to go from the functional prototype to the system in production? How much do I have to audit? How much do I have to review? Can I trust what was generated for an enterprise system?

The
re, GeneXus has an important construction difference.

Generative AI with Enterprise Governance

At GeneXus we do not ignore the benefits of generative AI. In fact, we built Glob.AI OS to be able to leverage it better in enterprise environments, and we built GeneXus Next on top of this platform.

Today, GeneXus (through GeneXus for Agents and other assistants such as Nexa) can use AI to assist, accelerate, interpret intent, and help build the model – but code generation does not depend on an LLM that generates different things in each interaction. Generation starts from a formal KB and deterministic generators. That means the generated code is not a different probabilistic occurrence in each interaction, but the consistent result of a model, proven patterns, and a platform designed precisely to produce executable systems from structured knowledge.

Obviously, testing is still required. You need to validate business rules, user experience, integrations, and functional behavior – but the nature and scale of the problem changes: you are not auditing a mass of code freely generated by an agent, but verifying a system produced from a governed source of truth.

The
Economic Factor of Coding Agents

Generative AI is very expensive. For individual developers or small projects, and for now, that cost is subsidized by the platforms.

The
re are plans that give you a large number of tokens at a reasonable cost. But in enterprise environments the situation changes: plans, prices, controls, and restrictions are not the same. And when we talk about Enterprise, we are not only talking about the development team. We are talking about companies where tens or hundreds of people start using AI to design, analyze, document, automate, review, generate, and operate processes. In an organization of 150 or more people, the cost of AI stops being anecdotal.

In that context, a strategy based solely on generative generation can become expensive, difficult to govern, and difficult to scale.

Why the Combination Matters More Than Substitution

Today the question we ask ourselves internally is: What happens when we combine the productivity of modern agents with a governed and declarative GeneXus knowledge base, generative AI, and deterministic generation?

We believe that is where the opportunity and difference for GeneXus lies in the enterprise environment.

GeneXus already demonstrated for decades that the path for creating enterprise systems goes through raising the level of abstraction: not writing everything by hand, but modeling knowledge and generating systems from that knowledge.
The
market has now reached a similar conclusion, though by a different path: code will be generated more and more.

It arrived by a different path, but it arrived: code will be generated more and more; the difference will be in where it is generated from, how it is governed, and how much it costs to verify it.

Now, with Glob.AI OS for governance, GeneXus for Agents for agentic interaction with GeneXus knowledge bases, and with GeneXus Next (and even GeneXus 18), we are working to demonstrate that the path of combining artificial intelligences gives us the best results.

In a talk, Gaston Milano (CTO Glob.AI OS) spoke precisely about this – the difference is no longer only in the model, but in the entire engineering harness surrounding the model: the tools, memory, filesystem, databases, execution loops, control mechanisms, and the capabilities given to the agent.

And there a key word appears to understand where the industry is heading: neuro-symbolic. That is, generative AI creating text or code, combined with engineering, rules, tools, memory, control mechanisms, and knowledge structures that constrain, guide, and validate what the model can do.

At GeneXus, that idea has a very concrete translation. We do not start from loose code, but from knowledge formalized in a knowledge base. And we do not generate solely through a probabilistic output of an LLM, but by combining generative AI with a symbolic/deterministic layer based on rules, inference, and proven generators. In current terms, we could say that GeneXus is evolving toward a neuro-symbolic architecture applied to software development: generative agents to capture intent and operate, and a deterministic knowledge base to govern, validate, and generate sustainable, auditable, and evolvable enterprise systems.

This is our focus now: showing what can be built when agents do not work only on code files, but on an enterprise knowledge base, and when generation does not depend solely on probabilistic AI, but on governed models and deterministic generation.

To work in that hybrid world, using the power of coding agents and GeneXus at the same time, is why we created GeneXus for Agents.