The Agent Surface
The new default interface for work
We are entering the “decade of agents”1 but agents remain a spiritual entity to most. It can mean everything from a chatbot, copilot, sidebar etc. I want to concretize a viewpoint on how agents should be approached and presented to the masses.
A surface of work is the place where work lives, plus the primitives and conventions for getting that work done. Knowledge work primarily lives in documents, spreadsheets, and slides—intuitive, but passive. SaaS tools shifted work into specialized products by encoding workflows directly into software. Agents introduce the next surface of work.
Agents are a surface defined by three traits: natural language, autonomous intelligence, and generative capacity. These characteristics make it a super-tool compared to the previous two paradigms by combining their strengths. The natural language interface enables the intuitiveness of docs, while the on-demand generative capacity gives users the efficiency gains of specialization. The cherry on top is the autonomy to delegate -- you don’t have to do the work! Rather than having surfaces built for human actions, the surface of work becomes the conversation and the agent’s state: you express intent in natural language and the agent takes on the actions autonomously.
The agent surface isn’t obvious yet because AI’s most-used interface, ChatGPT, is mostly treated as a companion to help you with work, not the place where the work itself lives. That’s a key distinction from something like Google Docs, which is the canonical surface for a huge amount of knowledge work2. Agents will become this default surface where work is done3. I expect it to be as commonplace to create agents as creating a document in the coming decade4.
To explain the concept more concretely, it is a useful exercise to take a use case and reframe it into an agent. Travel planning is a good example. Today it lives in a mix of docs and spreadsheets (itineraries, budgets) plus SaaS tools like Expedia or Kayak that still leave you doing the coordination. In an agent-first world you’d create a new agent for the trip, brief it once on constraints and preferences, and it handles the rest. You talk to it throughout the lifecycle:
“find a red-eye under $500”
“optimize this day around the museum”
“pull my flight confirmation from email and add it to the itinerary”
“we missed the train, replan the afternoon”
“text John the updated plan and ask if he wants to join dinner”
“split the costs and send everyone a summary”
This exercise has reliably proven that most use cases can be expressed as an agent, and that the agent version is better. The intuition is that the shift from traditional software to agents turns products into services. For example, an executive assistant is the luxury version of self-serve tools like calendars, inboxes, and task lists. They take initiative, track details, and handle follow-through without constant direction. Agents democratize that level of service for any use case, with the added advantage of scale and consistency.
Aside: Consumer vibe-coding apps are almost too unambitious5 as they keep users in the pre-agent software era. There is no need for infinitely more SaaS tools, even if they can be built quickly on-demand. The focus should be on giving consumers the ability to easily build powerful agents. In-conversation generative UI will be the layer where on-demand software should be produced, acting as an interface between the human and AI system, not the product itself.
To satisfy most use cases I have found the core primitives6 to be:
frontier multimodal model
long context + memory
computer (filesystem + browser)
connectors
generative UI
composability7
Most agentic products currently do a subset of the above (which ends up being enough to satisfy the target vertical). A feature-complete agent surface will make building an agent with the above components intuitive to consumers and generalize across verticals.
Simply put, the agent surface will be the “computing 2.0” version of the Microsoft Office suite.
Coined by Andrej Karpathy on the Dwarkesh Podcast:
An easy litmus test is to ask yourself what do I share? I will share my google docs for collaborative work, not quite chats ... yet. Agents will become the unit of sharing.
Two ratios to watch as adoption spreads (especially in the enterprise):
Ratio of humans to agents
Ratio of agents to sub-agents
I would go so far as to say we create more agents than documents. An agent is a super-set of a document as it can also complete the intentions derived from the document. With generative UI, the agent can present you with a document editor if it is the best UX for the task at hand.
This isn’t to say vibe-coding products are useless. Coding products will continue to replace human-written code. They will effectively become autonomous software engineers. I am arguing the great agent product for the average consumer should not build just in time software, but rather agents.
