What It Means to Get Headless

to get: (1) Do something (2) / get command

Software engineering is going through its largest shift in the past decades with the arrival of agentic workflows. Tools like Claude Code now author up to XX% of requests to GitHub [TODO: add source], have caused a major market meltdown (the SaaSpocalypse), brought Anthropic to hitting a $30B run rate and made developers rethink their jobs.

This explosion in agentic workflows is extending to other fields and leading an enormous change in how software is being used and who uses it. Nothing illustrates this shift more poignantly than Ramp’s recent announcement to ship a CLI (command line interface) for users of their expense management software.

All modern expense management software combines an intuitive user interface with a linked credit card. Card expenses show up as transactions in a web interface. You then upload receipts and match them to the corresponding transaction (illustration of process here), through drag and drop, confirming automatic matching or uploading receipts individually for each transaction. Perfecting this workflow process and pairing it to sophisticated features such as spend monitoring gave rise to several $B+ expense management vendors like Ramp, Brex or Expensify.

A few weeks ago, Ramp shipped a custom CLI where an agent would complete the entire workflow. The agent figured out that a receipt was missing, found it on the user desktop and uploaded it. Ramp did the matching. A human directed this process through the Terminal but the GUI was bypassed entirely:

Ramp’s agent-driven expense workflow (via @RampLabs)

The user interface is no longer graphic — it’s based on text, machine-readable data and natural language. The target user is an agent. Not a human.

The headless economy

Beyond the GUI

Most businesses are not ready for agents becoming consumers of digital goods and services in their own right. The way agents make consumption choices will be very different from us and products they prefer look nothing like the ones most software businesses have been building. These new patterns of consumption will result in the rise of the headless economy.

The headless economy is the emerging market in which agents discover, evaluate, access, pay for, and consume digital goods and services through machine-native interfaces.¹ Headless software does not rely on a “head” or graphical user interface (GUI). It exposes API endpoints that allow the user to access data, manipulating it in whatever form they want to. Pre-agentic, the term was often used to describe “headless content management systems” — systems that allow you to surface content in a structured manner, such as this blog, but without supplying a GUI as a service.

The shift we’re talking about with the headless economy is not new. APIs have existed for decades and API-first businesses are not uncommon. What’s changed now is the actor. For most of software history, the entity navigating, selecting, integrating and paying for a product was a human. And using it — we click, log in, drag, match, upload, filter, select. If there was an API, we would implement it in deterministic fashion to power other GUIs. The GUI was the product.

Agents do not need a GUI. They need access to structured data. And in the headless economy, that will happen through APIs, MCP servers, CLIs. Headless implies that a human-legible GUI is not the primary consumption layer.

When the primary consumer of a good changes, the product needs to catch up. The headless economy is specifically about how software gets consumed and who or what is doing the consuming — with the following implications:

1. The value of investing in human-centric interfaces will decrease, unless they marry human-native and machine-native
2. The value of businesses building infrastructure and abstracting away complexity from agents will increase

Stripe still needs to build world-class financial infrastructure. YipitData still needs to create proprietary datasets. Ramp still needs to provide world-class receipt → transaction matching and issue credit cards. The value doesn’t disappear. The interface to that value changes.

And it will give rise to a completely new type of business, one that doesn’t need to cater to humans at all. And others are noticing too:

“The biggest opportunity in agentic commerce (…) is building headless merchants. The next generation of merchants won’t have storefronts. They’ll have endpoints.” — Noah Levine, a16z

“Agents don’t read your ad. They query a registry, get structured results, and pick the best option in milliseconds. Your pricing page becomes a machine-readable header. Your sales funnel collapses into three HTTP calls: discover, authenticate, buy.” — Simon Taylor, Fintech Brainfood

Why now

What unlocked this

Agents became capable of completing complex, multi-step tasks — not just generating text. What made that possible was giving them access to tooling: APIs, CLIs, MCP servers, structured data feeds as well as the ability to chain these toolings together in tight reasoning loops. One implication is that once an agent can call a tool, it becomes a user of that tool. And users have needs — discovery, authentication, pricing, reliability, trust.

This in turn unlocks a lot of new opportunities like providing pre-vetted and authenticated data to agents at cheap prices. When faced with a choice between researching and computing an answer by itself or purchasing that information from a vendor, given that agents are natural optimizers (that’s how their training works!), they would be able to pick the most cost efficient choice.

Simon Taylor puts it nicely in the following table in “The Intention Layer”:

	Agent self-computes	Buys from a specialist
Cost	$0.10 – 0.50	$0.01 – 0.02
Speed	10 – 25 seconds	< 200 milliseconds
Advantage	—	7-50x cheaper, 50-100x faster

The shape of what’s coming

Agents don’t all consume in the same way. There are three distinct patterns, which we will cover in future posts:

• Agent-as-consumer: The agent directly consumes the capability or output as part of its own workflow.
• Agent-as-intermediary: The agent acquires a good or service on behalf of a human or organization, but is not the ultimate consumer of that good.
• Hybrid: The agent consumes intermediate goods or capabilities in order to generate an output for a human or another system.

The journey an agent takes from discovering a capability to repeatedly consuming it follows a (for today!) similar-ish funnel than the human customer journey you’re used to thinking about:

1. Discover
2. Select
3. Access
4. Consume
5. Pay

/get-headless

Why this blog

My first job was in taking terabytes of datasets on publicly traded companies and distilling them down into 5 bullet points. This enabled investors in said companies to make well-informed bets on their stock prices. The basics behind it — structuring, slicing, reducing, querying, vetting complex sets of data into single verifiable facts — will be what power the headless economy.

/get-headless aims to be a knowledge hub for everything around this new market. I personally am extremely intrigued by the shapes of this market and all the opportunities it brings. Some key questions I want to answer:

• What businesses are being built specifically for agent demand?
• How do incumbents adapt existing products for machine-native consumption?
• What is the infrastructure enabling it? APIs, CLIs, SDKs, MCP servers, agent protocols?
• How is pricing and access models for machine consumers going to change?
• How will discovery, selection, and distribution work when the customer is software?
• How to think about trust, identity, and authorization in agentic systems

And we’re early. The businesses that will define this market are being built now, and most of them don’t have a name yet. This is their home.

Footnotes

1. Some call these “machine” or “agent-to-agent (A2A)” markets. ↩