G5 Labs wants natural-language intent to replace code as source

What G5 is building

G5 is a secure cloud web platform that turns business requirements, architecture decisions, policies and other human intent into a system ontology: a structured semantic graph describing the goals of an entire system and the relationships between people and AI agents.

Kraska told VentureBeat that G5’s hypothesis is to make natural language a new kind of source code. Natural language, augmented with the structure of a system ontology, becomes the source representation; Python, Rust or another programming language is generated from it.

That sounds close to specification-driven development: describe what a program should do, give the specification to an AI agent and let it produce the implementation. G5 is claiming a larger role, however. Its semantic representation is meant to:

Kraska calls this a new level of abstraction rather than improved prompting. In G5’s model, natural-language intent can be compiled, merged, compared and governed by rules. The intended audience is therefore not only developers, but product owners, compliance specialists and other people responsible for what software does.

The company’s core mechanism is bidirectional. Existing software can be lifted from source code into an ontology. People and AI agents can then reason over that representation, while G5 pushes approved changes back down into ordinary programming languages and frameworks.

G5 says its bidirectional compiler can:

Each ontology node is written in natural language and linked to the code that implements it. Chief operating officer David Chang, who previously worked at Google and led the global go-to-market function for Amazon’s Nova AI models, said the system covers:

G5 says it continuously checks the ontology for consistency and the code for conformity with the ontology. Its website describes the relationship simply: change the intent and the code is rewritten; change the code and the ontology learns. In a company demonstration, generated source code included tracing metadata pointing to a specific ontology concept and version.

That traceability matters if AI-generated code grows beyond the amount humans can inspect line by line. Kraska said one G5 engineer submitted a 300,000-line change in a week. He argued that new validation mechanisms are already needed at 100,000 lines because older review methods no longer cope.

G5 also demonstrated a “change graph” connecting human chats and documents to ontology concepts, architecture components and individual source files. Other interfaces showed ontology and implementation checks, implementation progress, test coverage, and separate stages for merge-request creation, rebasing, approval and merging.

The practical proposition is semantic review instead of text-only review. If two developers or agents modify a banking application, one might require a red button while the other makes no statement about color and happens to generate blue code. A normal source-code merge could produce a conflict. At the semantic level, G5 could theoretically recognize that only one side made a requirement and merge the changes automatically.

If two participants specify incompatible authentication requirements, the system should identify a real semantic conflict and send it to a person.

Modernizing systems and replacing SaaS

G5’s early customers are using it for two related jobs: modernizing existing software and replacing expensive third-party SaaS with applications the customer owns and controls.

Kraska said roughly 90% of the organizations working with G5 are in financial services. He described the projects as early and did not name the customers because of confidentiality agreements. VentureBeat saw a G5 demonstration based on a real modernization project involving a large enterprise system, but the customer remained anonymous.

The company’s modernization workflow starts by recovering the meaning of a legacy application rather than translating its source code directly into another language. G5 argues that conventional AI migration carries old assumptions and technical problems into the new system, including decisions retained only because of historical project constraints.

Its proposed workflow is to:

G5 says ontology comparisons in a financial-sector modernization project exposed structural differences that ordinary code migration would have missed.

The second use case is replacing SaaS. G5 says companies could build applications around their own workflows instead of continuing to pay for general-purpose software, with potential savings of millions of dollars. Kraska told VentureBeat that as implementation becomes cheaper, highly customized software will increasingly compete with high-margin SaaS products.

That is a much broader claim than “AI writes code faster.” It suggests that the economic boundary between buying software and building it could move if companies can describe their own processes well enough and maintain the resulting systems.

The evidence is still supplied by G5 itself. The company says its product is deployed in several regulated industries and that projects have lifted “millions of lines” of legacy code into semantic ontologies. VentureBeat has not independently verified those deployment or scale claims.

G5 is an orchestration layer, not a model company

G5 is not building another frontier coding model. It sits above models and coding tools, assigning work to systems such as Claude Code and Codex and choosing different models for different jobs.

The platform is not tied to one model. Customers can select coding tools and underlying models, while G5 can route implementation work to Claude Code or Codex and use cheaper systems for less demanding tasks. That could appeal to companies that do not want their development-control layer tied to one model provider.

G5 says it can:

This makes G5 more than a semantic registry. If the system works as described, it also decides how work should be divided, what constraints agents should follow and how their output should be reconciled.

The platform includes cost checks as well. Kraska described cases in which a small natural-language request actually implied an architectural rewrite. G5 can estimate the expected cost in advance and request further approval before implementation begins.

Governance is central to the pitch. Kraska argues that responsibility does not disappear because an AI agent created the change. Developers, in his view, will remain but become managers of virtual employees: agents that handle more implementation while humans make high-responsibility decisions about architecture, operations and related matters.

G5 connects changes to approval processes and policies so a company can record:

My read is that this is where G5 has a more defensible position than another natural-language coding interface. The company is trying to own the control plane around agents, not win the model race. But it also means customers would be adding another critical layer to the software stack, rather than eliminating spending on underlying models and coding tools.

The old idea in a crowded new market

The premise that a durable model should outlive generated code predates generative AI. In 2001, the Object Management Group made Model Driven Architecture a foundation of its standards. MDA starts with a platform-independent model of business functions and application behavior, converts it into platform-specific models and implementations, and anticipates regenerating code when the underlying technology changes.

AveriSource offers another comparison. Its tools analyze legacy applications, extract business rules and data structures into an intermediate domain model, and use that model to create implementations in Java, C# and microservices architectures. Its current Reimagine product also uses structured business-function specifications as inputs for AI-oriented direct development.

That does not make AveriSource equivalent to G5. G5 claims to cover current development, semantic merging, natural-language intent and enterprise governance. But extracting stable meaning from old code and regenerating a modern implementation is not unique to G5.

The competitive landscape now includes products that already make specifications, agent context or governance part of development:

G5’s claimed distinction is the persistent semantic graph and its bidirectional relationship with implementation. It also wants to preserve human intent separately from agent decisions, identify semantic conflicts, apply policies across the graph and coordinate multiple agents.

That distinction is technically meaningful if it holds at enterprise scale. It is also harder than producing specifications in Markdown.

The commercial alternatives make the comparison concrete:

G5 has not published prices. The company currently offers early access rather than self-service purchase, reinforcing that it is being sold as an enterprise engagement rather than a tool an individual developer can deploy immediately.

The enterprise test is portability

G5 says its standard delivery is a cloud environment accessed with customer credentials. It also says it can work with banks and other organizations with stricter infrastructure requirements, keep data with the customer and operate within existing agreements with model providers, including no-data-retention modes where those providers support them.

Customers can upload an existing system and build an initial ontology themselves or with G5. They then modify the representation and regenerate the application.

The unresolved cost is larger than a platform subscription. Buyers will also need to account for:

For a serious pilot, I would want to see G5 work on a real application with accumulated historical code, not a clean greenfield demo. The important tests are whether it can:

Several operational questions remain publicly unanswered:

That last question is the most important. If the ontology becomes the durable record of business intent while conventional implementations are repeatedly regenerated, the portability of the semantic asset may matter more than the portability of the generated source code.

Investor and former Google employee Jeff Dean called G5 a serious, well-designed attempt to raise the abstraction level of programming again. Battery Ventures partner and former MongoDB CEO Max Schireson went further, arguing that AI will remove the need to write software in Java, C, Python or Rust and that English will become the new development language. He described G5 as the missing tooling for building a reliable development process around that language.

I think the immediate enterprise question is narrower and more difficult. It is not whether English replaces Java. It is whether an always-changing AI-generated codebase can be governed by a semantic layer without that layer becoming another stale artifact—and whether customers can take that layer with them if G5 does not remain the place where their software lives.