You're using an "Internet Explorer" web browser that's no longer supported.
This website will not function correctly when using when using Internet Explorer,
so please upgrade to a modern web browser, like Edge, Chrome, or Firefox.
Led by Instantiations CEO Seth Berman, the session demonstrated how AI agents can interact with VAST Platform applications, use existing Smalltalk operations, and support development workflows. It also explored the engineering considerations involved in bringing these capabilities into production environments.
Through a series of demonstrations, Seth showed how VAST applications can participate in agentic workflows without replacing the business logic they already contain.
Building on the knowledge already in your applications
Enterprise applications contain more than code. They represent years of business decisions, calculations, validation rules, integrations, and operational knowledge.
When introducing AI, the goal does not have to be recreating that knowledge in a language model.
Instead, AI can provide new ways to interact with existing application capabilities while the application continues to enforce its business rules.
To demonstrate this approach, Seth used the VAST AI Assistant to create a small quote calculator. The application accepted a quantity, unit price, and customer discount, then calculated the resulting total.
The example was intentionally simple: five units at $100 each, with a 10% discount, produces a $450 total.
The more important detail was how the calculation worked.
The language model interpreted the user's request, but the actual calculation remained in a Smalltalk business object. That same calculation could then be used by a native interface, an AI agent, and a Seaside web application.
This separation provides a practical starting point for adding AI capabilities to existing applications.
Introducing agentic workflows with VAST ADK
A significant portion of the webinar focused on the VAST Agent Development Kit (ADK), Instantiations' native Smalltalk implementation of Google's ADK programming model.
A model can interpret language and help determine what action to take. An agent framework provides the surrounding capabilities needed to coordinate that work within an application.
The VAST ADK development work includes:
Connections to multiple AI model providers, including Gemini, OpenAI, Anthropic, Ollama, and Azure OpenAI.
Tools that invoke Smalltalk operations or connect to external capabilities.
Sequential, parallel, and looping agent workflows.
Session state, memory, and supporting artifacts.
Execution controls, confirmations, and human approval.
Evaluation and tracing to help examine agent behavior.
These capabilities become important when an AI interaction moves beyond answering questions and begins participating in business processes.
For example, an agent might retrieve information, call an existing calculation, ask for confirmation, and continue to another step.
The application must still determine which operations are available, what inputs are acceptable, and which actions require additional review.
Connecting systems through MCP
The webinar also explored the Model Context Protocol (MCP) and how it can connect AI tools with existing systems.
In one demonstration, the VAST AI Assistant used an MCP connection to retrieve relevant VAST documentation while investigating a Smalltalk method.
Rather than relying solely on the model's existing knowledge, the assistant could access documentation and provide information that the developer could inspect.
MCP also creates opportunities in the other direction.
VAST can expose defined capabilities through an MCP server so that compatible external AI clients can discover and invoke those operations.
The distinction is important: an agent can call a local Smalltalk operation directly through ADK, or use MCP when an external connection is appropriate.
The choice depends on the application's architecture, the capabilities required, and how access should be controlled.
The VAST Platform MCP SDK is available as an open-source project for developers interested in exploring these integrations.
The demonstrations also explored how developers and users might interact with AI-enabled applications.
Using Gemini Live connected to the VAST development environment, Seth demonstrated a spoken conversation in which the assistant inspected application code, generated architecture and sequence diagrams, and exported a diagram to PDF.
The value extended beyond voice interaction. The assistant was connected to the actual development environment and could use tools to investigate the application.
Another demonstration introduced A2UI, an approach for allowing agents to produce structured interactive interfaces.
Rather than returning only text, an agent could provide a form, supporting information, or interactive controls that the host application validates and renders.
These capabilities suggest additional ways for people to interact with application functionality without requiring every interaction to begin with a traditional interface.
Human approval and governed execution
As agents gain access to application operations, an important question emerges: what should they be allowed to do?
The webinar included a demonstration of an operation configured to require human approval before writing a file.
When the request was rejected, the operation did not execute. When it was approved, the file was created and its contents could be inspected.
VAST Python
The session also previewed VAST Python and governed runtime research.
In that demonstration, Python executed inside VAST with defined access and resource limits. An operation outside the permitted interface was rejected, while an infinite loop was stopped after exceeding its execution budget.
This work remains in early-stage development, but it illustrates an important engineering direction: defining and enforcing boundaries around what agent-generated code can access and execute.
AI development still requires engineering judgment
One of the broader lessons from the webinar was that faster code generation does not eliminate the work required to deliver reliable software.
AI may help developers create an application more quickly, but testing, security checks, packaging, documentation, deployment, and ongoing support remain essential.
It also changes the questions engineers need to ask.
What is the application supposed to accomplish? What constraints matter? How do we know the result is correct? What evidence is needed before it reaches production?
These questions become increasingly important as AI takes on more implementation work.
The objective is not simply to generate more code. It is to improve outcomes for the people using and maintaining the application.
Where should VAST customers begin?
For organizations exploring agentic AI, the webinar offered a practical recommendation: start with one existing operation.
Choose a business function that is well understood and produces a result that can be independently verified.
That might be a calculation, an information lookup, a validation rule, or another established application capability.
From there:
Identify the operation and its dependencies.
Determine who should be able to use it and what it can change.
Expose a defined tool interface.
Test the interaction with realistic requests.
Evaluate the results with the people responsible for the business process.
Review the controls, testing, and delivery requirements before expanding the implementation.
This approach provides a manageable way to explore AI capabilities while building on the systems and expertise already in place.
Continue exploring VAST and AI
Instantiations is developing AI capabilities for VAST Platform while working with customers to explore practical applications of the technology.
The VAST AI Assistant and VAST Platform MCP SDK are available as open-source projects. Additional capabilities demonstrated during the webinar, including VAST ADK, A2UI, and VAST Python, remain under development.
Watch the complete webinar to see the demonstrations and hear the technical questions raised by the Smalltalk community.
Interested in exploring where AI could help your existing VAST applications? Contact Instantiations at info@instantiations.com to discuss a specific workflow or application opportunity.