As Agents Get Stronger, Why Do Enterprises Need a Runtime Even More?

Over the past year, it has felt as though the software industry knocked a hole through the wall between writing code and doing work. Claude Code, Codex, and CLI agents no longer stop at answering questions. They can read files, run commands, call tools, modify projects, and carry out extended sequences of complex tasks.

It is natural to ask: If agents are becoming so capable, do enterprises simply need to keep adopting stronger models and agents?

We are increasingly convinced the answer is no.

Stronger agents expose another problem: being able to do something is only the beginning. Enterprises need to organize those capabilities, reuse them, trace them, and keep them running.

That is what ZGI has been working on.

Our focus is the practical problems agents encounter in real operations: bringing models, knowledge, data, Skills, workflows, and runtime governance into one system.

If agents are already capable, why do we still need workflows?

This is an increasingly common question.

If an agent can assess a situation, select tools, and execute tasks, why draw a workflow? Why define nodes, conditions, inputs, and outputs? Will stronger models eventually make all of this obsolete?

Think of an agent as a highly capable person, and the answer becomes intuitive.

Even a very smart employee does not remove a company's need for processes.

Sales uses a CRM, engineering uses issues and CI/CD, finance has approval rules, and support has escalation procedures. Organizations need these processes to make work stable, repeatable, and deliverable.

An agent addresses how an individual task is completed. A workflow addresses how those capabilities are organized to work over time.

An agent can assess whether a customer email signals strong interest. A business still needs to specify where that lead is recorded, who is notified, when execution continues automatically, when a salesperson must confirm, whether failures are retried, and how many consecutive failures should stop the process.

Those requirements do not disappear as models improve.

In fact, stronger execution capabilities make clear boundaries more important.

This is how we understand workflows: they help turn agent capabilities into organizational capabilities.

Conceptual illustration: agents and workflows have distinct roles within a shared runtime.

From “it managed it this time” to “we can give it this job from now on”

One exciting aspect of agents is their ability to quickly address needs that previously did not justify dedicated development.

Traditional software development has a practical threshold. A niche requirement may need only a few hundred lines of code and still never make the engineering roadmap.

Operations wants a daily report combining several channels. Sales wants customer replies assessed and synchronized to the CRM. A content team wants a daily scan of industry sites and communities, with genuinely relevant items collected into a briefing.

These requests used to occupy an awkward middle ground.

Manual work was repetitive and wasteful, while dedicated development offered an uncertain return.

Agents change the economics of that work.

They can generate scripts, understand unstructured material, call different tools, and choose the next step based on intermediate results.

But running successfully once does not mean a team can rely on the process indefinitely.

What happens if the model changes tomorrow? If an API times out? If the data format changes? If the third step fails? If the employee who built it leaves?

ZGI focuses on that next stage: turning an improvised success into a durable capability.

Models can change while workflows remain. Execution can improve while Skills are reused. Knowledge bases and data sources can be updated without rebuilding the business logic from scratch.

We want capabilities that can remain part of an enterprise over time.

Skills mean more than installing extra plugins

Skills have attracted a great deal of attention.

They are often described like apps on a phone: add an Excel Skill today, a database Skill tomorrow, and a drawing Skill after that.

For an enterprise, a more valuable interpretation is capturing capabilities scattered across people, scripts, and prompts as reusable assets.

A daily sales report Skill might be straightforward: query a few sources, aggregate data, ask a model to analyze anomalies, and produce a standard output.

Once packaged, however, it no longer depends on one employee's personal know-how.

Sales can call it. An executive's agent can call it. A weekly business review workflow can call it.

Customer background checks, contract summaries, and internal knowledge retrieval can likewise become part of a company's own library of AI capabilities.

That is why Skills matter to us.

Models will keep improving. The capabilities an enterprise accumulates should not need rebuilding alongside them.

Enterprises need more than a smarter agent

Imagine an agent that can already read files, call databases, generate documents, and execute code.

That sounds promising.

An enterprise technology leader will still ask which model it used, why a run consumed so many tokens, which data it accessed, and why a particular task failed.

Was the error in the model's judgment or a tool call? Who is allowed to run the agent? If it takes an incorrect action, can the sequence be traced?

These questions are easy to overlook during a demo. In production, they become basic requirements.

ZGI therefore pays attention to the execution process itself, not just the number of things an agent can do.

Model routing, runtime logs, node state, inputs and outputs, token use, costs, permissions, and errors may be less eye-catching than a new agent demo. They determine whether a business can trust the system with real operations.

In this sense, Agent Runtime brings familiar software principles into AI: observability, control, recovery, and auditability.

AI can become more autonomous without production systems becoming harder to explain.

Model Gateway is more than connecting additional models

It is increasingly common for one enterprise to use several models.

Claude may suit some long-document tasks, GPT some reasoning tasks, DeepSeek some cost-sensitive batches, and private models certain sensitive internal processes.

The question shifts from choosing a model to managing models.

If every agent stores its own API keys, maintains its own configuration, and calculates its own costs, more agents mean a greater administrative burden.

ZGI's Model Gateway aims to make models centrally managed enterprise resources, rather than connection details scattered across dozens of projects.

Providers, call strategies, and prices can change. Applications and workflows should not need repeated restructuring because of those changes.

We want to separate model capabilities from business logic as far as practical.

Faster model development makes that separation more valuable.

Source access and self-hosting are more than technical preferences

When AI helps an employee draft a few emails, its deployment location may not seem critical.

Once agents access internal knowledge, customer information, databases, and real business systems, control becomes a practical concern.

Teams need to understand how the system operates, where data flows, which services are accessible, and which capabilities must stay inside their network.

That is why ZGI continues to make its source available and supports self-hosting.

The underlying principle is simple: the deeper AI enters a business, the more that business should understand and control its infrastructure.

Teams can choose their models, connect their databases, run on their infrastructure, and extend the system around internal security and compliance requirements.

We believe this will be difficult to avoid as enterprises use agents over the long term.

A more realistic future: many agents working together

A common vision of AI's future centers on one all-powerful assistant that handles everything.

We see enterprise environments somewhat differently.

Organizations are more likely to have many agents. Sales agents understand customers and CRM. Engineering agents work with code and documentation. Finance agents focus on data and processes. Operations agents continually process content, metrics, and external information.

Their models, permissions, knowledge, and Skills differ.

Some tasks can be delegated entirely. Others require workflows that coordinate several agents. Still others must move back and forth between people and agents.

The valuable capability will be organizing agents, Skills, data, and people into a system that can operate over time.

That is ZGI's long-term understanding of Agent Runtime.

We do not expect agents to eliminate workflows or stronger models to make infrastructure unnecessary. Stronger agents make organizational coordination more important.

Models handle improvisation. Skills provide capabilities. Workflows coordinate work. Runtime keeps it all operating.

There is still a substantial distance between “AI completed this for me once” and “we can reliably entrust this work to AI from now on.”

ZGI aims to shorten that distance.