Ir directamente al contenido principal

Headless software raises the bar for what we build

AI agents have changed how users access software. The ripple effect is that product and engineering teams need to deliver insights, actions, and outcomes through experiences they don’t control.


Jackson Wolfe headshot

Jackson Wolfe

Group Product Manager at Zendesk

Última actualización el 19 de noviembre de 2026

Graphic with the text Building beyond the interface

For a long time, we've built software around the assumption that people will interact with it through an interface. Navigation, dashboards, configuration screens, and workflows. Product and engineering teams invest significant time making it easier for people to tell software what to do.

But what happens when the person benefiting from the software no longer needs to operate it? A customer might complete a transaction through a mobile app, access functionality through a partner integration, or trigger a workflow from another system entirely. The underlying software still does the work, even when nobody opens the application it belongs to. 

Headless architecture separates a product’s capabilities from the interface used to access them. AI agents make that separation more consequential: another system can now decide how to use our capabilities, interpret their outputs, and present the results to the customer.

The same capability, a different experience

Working on ways to access software through a terminal or coding agent has sharpened how I think about headless. These experiences let customers ask questions and take action outside the product’s own interface. The software still does the work, while the terminal or coding agent shapes how that work is presented.

When someone uses software directly in a terminal, we can shape how it communicates: how progress appears, how tool calls are surfaced, and how actions and results are explained. When a coding agent uses those same capabilities, it interprets the output and decides what to show the user. The capability is still doing the work, but another system controls how that work is presented.

That creates a different product challenge. How do we communicate what our agent is doing when we don’t control the display? How do we make actions traceable when another agent might summarize the result? What information needs to travel with an output so it can be used correctly in a different context?

The agentic offering behind the interface becomes critical. Its product knowledge, supported actions, and ability to explain what happened need to remain useful across both experiences.

I find this freeing. It pushes us to get specific about the value our product provides: the insights we can surface, the actions we can support, and the outcomes we can achieve. That value should be available wherever the customer needs it.

What changes when software moves beyond the interface?

Building for this environment means making our capabilities understandable to other systems, dependable wherever they are invoked, and accountable for the results they deliver.

I see three priorities for building software in this environment.

1. Build capabilities that can be reused and extended

Building for multiple surfaces starts with understanding the underlying customer need. A customer might ask for a dashboard, a report, or a configuration screen. Each is a way to access information or accomplish something. Getting to that underlying need helps us build capabilities that can be reused and extended across experiences.

AI agents introduce another consideration: discoverability. As I explored in my previous article on Skill-Driven Development, agents need to understand which capabilities are available, when to use them, and how to invoke them correctly.

APIs remain foundational. MCP, agent-to-agent communication, and agentic interfaces like terminals and coding agents can help expose tools and expertise in ways other systems can discover and use. We use Agent-to-Agent (A2A), but the broader principle applies across frameworks: make it easy to obtain relevant knowledge and take the right action.

The architectural challenge is deciding what stays consistent and what can adapt. Business rules and the meaning of an action should remain stable. The information returned can be selected and reshaped for the context in which it will be used.

A terminal or coding agent can provide progress summaries, tool activity, results, and suggested next steps as distinct pieces of information. That gives the receiving system something useful to work with, whether it is rendering a screen or deciding what an agent should do next.

2. Put reliability and control into the underlying system

When we control the interface, we can put a confirmation before an action or an explanation beside an error. When another system invokes our capabilities, those protections need to exist in the product itself.

I think about safeguards in two categories. Some protect the product: enforcing permissions, defending against prompt injection, and preventing unauthorized actions. Those are our responsibility as builders.

Others express the customer’s business policies. A customer might specify that information about an upcoming release cannot be shared with a particular audience. The customer should configure that policy, and the product should enforce it. Both categories need to apply before information is returned or actions are taken through an agentic interface.

The same responsibility applies to errors. We may not control how another system presents a failure, but we control the accuracy and usefulness of the information we return. Did the action fail before making any changes? Did it partially complete? Is retrying appropriate? Those distinctions matter when another agent is deciding what to do next.

Traceability, logs, permissions, and safety become central to the offering. The product should preserve a reliable record of what it did and provide enough context for another system to explain it. A summary on an external surface should still be traceable to the underlying actions and results.

The most useful product insight comes from combining usage and outcome data.

3. Measure the value delivered, wherever it happens

We've spent years measuring product adoption through activity inside our products. As software becomes accessible through more channels, traditional measures of product engagement tell us less about the value customers receive. A capability might power a partner integration, an automated workflow, or another product without anyone opening the original application. 

We still need to measure usage: which capabilities are invoked, what inputs they receive, how long they take, whether human intervention is required, and what they cost. But those measures need to connect to the result the customer values. Token consumption tells us something about the resources used. It doesn’t, by itself, tell us whether the work was useful.

The right outcome depends on the business. Zendesk’s Outcome AI™ announcement makes this concrete: customers pay for eligible issues resolved and business work completed, with outcomes evaluated for quality and completeness. That creates a direct connection between the work the product performs and the value the customer pays for.

As we support more use cases, the outcome could also be a qualified lead or another business result. Each needs a definition the customer understands and the product can reliably instrument. We should be able to show what happened, why it qualifies as a successful outcome, and how it contributes to the customer’s goal. Successfully executing a tool call is only one step toward that result.

The most useful product insight comes from combining usage and outcome data. A capability that is frequently invoked but rarely achieves the intended result needs improvement. One that consistently delivers value but is rarely used may need better discoverability or access. Changes in inputs can reveal needs our existing capabilities don’t yet address.

Together, these signals help product teams decide what to improve, what to extend, and what to build next.

The product behind the interface has to earn its place

Headless also makes it easier to see where a product’s differentiation actually lives. When another system provides the interaction, the underlying product has to earn its place through the knowledge, capabilities, and results it contributes.

That connects to the deployment challenge I’ve written about before: customers receive value when software can be put to work in their businesses. Making our capabilities usable through experiences we don’t control expands where that can happen.

As my colleague Ben Mitchell explores in his article on designing for an interface that is everywhere, AI is changing where and how people experience software. Interfaces will still matter wherever they help people understand, decide, create, or enjoy the work. But the product behind them needs to stand on its own.

For me, that is the opportunity in headless. It forces us to be clear about what our software does for customers—and build something valuable enough that they want it available everywhere.

Jackson Wolfe headshot

Jackson Wolfe

Group Product Manager at Zendesk

Jackson Wolfe is a Group Product Manager, leading product development for Forethought AI Agents by Zendesk where he builds the agentic platform on which CX teams confidently build & deploy AI Agents to resolve customer queries and deliver delightful experiences to their users.


Residing in San Francisco, CA, Jackson is passionate about democratizing generative AI for CX teams, driving agentic innovation, and golf.