A company's first AI applications often emerge without a central plan.
Engineering builds a coding assistant. Operations creates a content agent. HR tries resume screening. Support creates a knowledge bot. Each team finds a model, writes a few prompts, and gets something working.
At first, it is exciting.
Three months later, the problems appear. Some teams use GPT, some Claude, and others DeepSeek. Several knowledge bases duplicate the same material. API keys live in different projects. Only one person understands a particular workflow. An agent suddenly consumes many tokens, and nobody can identify the step responsible.
AI applications are multiplying, while their management still resembles a collection of personal tools.
Many companies pass through this stage. The first question is whether they have AI. The second is whether those applications can become company capabilities.
ZGI focuses on that second question.
Does one enterprise really need dozens of agents?
It probably does.
Support and engineering solve different problems. HR and sales use different data. Finance operates under different permissions. It is difficult to imagine every employee relying on one universal AI.
The more realistic future includes many agents.
The problem is not their number, but whether they operate in isolation.
Three departments connecting to DeepSeek can mean three sets of API keys. Four teams using product information can mean four knowledge bases. The same Excel generation capability may be implemented repeatedly.
This resembles the growth of enterprise SaaS. Each tool can improve productivity on its own. As the number grows, shared identity, permissions, data, and administration become necessary.
Agents are going through a similar transition.
ZGI is therefore more than a way to create another agent. We think of it as an Agent Runtime Workspace where models, knowledge, data, agents, Skills, and workflows operate and are managed together.
First: Do not make every team reconnect the same models
The pace of model development suggests that few companies will use just one model indefinitely.
A task may suit Claude today and GPT tomorrow. A cost-sensitive batch may use DeepSeek. Sensitive internal work may require a private model.
ZGI brings models into a shared Model Gateway.
The purpose is not to collect the longest list of supported models. It is to avoid permanently coupling a business process to one model.
A workflow that has run for six months should not need to be rebuilt because the provider changes.
Models are resources. Business logic is the company's asset. Separating them as far as practical is part of making enterprise AI sustainable.
Second: Knowledge bases should not duplicate one another
HR uploads an employee handbook, then administration uploads it again. Sales builds a product knowledge base, and support creates another copy.
Six months later, nobody knows which is current.
We would rather treat knowledge bases as enterprise AI assets than as attachments to individual agents.
Company documents and knowledge can be managed together, with different agents accessing them according to permissions.
Live information does not all belong in a vector database. Orders, inventory, and customer records already exist in operational databases. Agents should read them through queries or Skills instead of depending on daily Excel uploads.
The difficulty of enterprise RAG extends beyond document chunking. Teams must decide what belongs in shared knowledge, what must be queried live, and who may see it.
Third: Package repeated capabilities as Skills
Suppose five agents need to generate Excel files.
Writing “Please generate an Excel file” separately into all five agents is a poor approach. A shared Excel generation Skill is more useful.
Database queries, report generation, chart creation, and internal API calls can likewise become reusable organizational Skills.
The immediate change may look small, but its long-term value grows with the number of agents.
If every new agent requires fresh data connections, prompts, and tools, maintenance costs rise with adoption.
A more scalable pattern is to add agents while reusing the capabilities underneath them.
That is why we see enterprise Skills becoming important assets.
Conceptual illustration: different agents reuse shared capabilities while retaining access and business boundaries.
Fourth: Capable agents still need established procedures
If models can plan tasks on their own, why retain workflows?
Because they address different needs.
Agents are good at uncertainty: analyze some material and decide what to investigate next.
Workflows carry established rules: who must approve a review, what spending threshold requires a pause, or how many consecutive failures require human intervention.
Real business combines both.
In ZGI, a workflow can connect models, knowledge bases, databases, Skills, conditions, and human steps. Agents can exercise judgment where it is appropriate, while established rules remain explicit.
Agents provide flexibility. Workflows define boundaries.
After launch, the questions change
When an agent first succeeds, people focus on the result: “It actually did it.”
By the thousandth run, priorities look different.
Why did it fail seventeen times yesterday? Which model costs the most? Why has a workflow slowed down? Can this employee use the finance agent? Did an incorrect answer originate in the model or the database?
ZGI's runtime layer therefore considers execution records, node inputs and outputs, model use, tokens, costs, and permissions.
These features may be less visually impressive than a demo. They are central to whether companies will put AI into production.
Software systems need more than the ability to run. They need to be observable, traceable, and controllable. Agents are no exception.
Why make the source available and support self-hosting?
For one or two small AI tools, cloud services are often the easiest path.
As agents begin accessing customer data, internal knowledge, databases, and business systems, more teams want to understand and control the full operating environment.
ZGI therefore makes its source available and supports self-hosting. Teams can run the components on their own infrastructure, connect internal models and services, and extend the system for their business.
The licensing boundary should be clear. ZGI currently uses the ZGI Community License. Personal, research, educational, and internal organizational use can be free. Certain commercial uses, including hosted multitenant and white-label offerings, require commercial authorization.
We do not want to describe that as unrestricted free commercial use.
For an enterprise, the more important question is whether core AI capabilities can be inspected, adapted, and deployed somewhere the company controls.
Which companies do not need ZGI yet?
This is worth saying directly.
A company that occasionally uses ChatGPT to write copy or summarize meetings does not need a full Agent Runtime. A mature SaaS product may also be simpler for one straightforward knowledge Q&A requirement.
ZGI is more relevant when several models, agents, or workflows are already in use; when AI must connect internal knowledge, databases, and systems; when teams want shared Skills; or when permissions, token use, logs, and private deployment become important.
Runtime becomes valuable as the question shifts from building one AI application to running the company's AI applications together.
Let the company's capabilities accumulate
AI applications may follow a path similar to SaaS. At first, companies ask whether they have the tools. Later, they have many. Eventually, the important question is how those tools are organized.
Agents are reaching that stage.
A company may eventually have dozens or hundreds of agents using different models for different tasks. Its knowledge, data, Skills, permissions, and business processes should still be able to accumulate over time.
Models can change. Agents can be rebuilt. The company's accumulated AI capabilities should not need to start from zero each time.
That is why ZGI is building an Agent Runtime: to help the AI already being created become a lasting part of what the company can do.
Top comments (0)