How much growth potential does Agent Native's infra have?

Agent is one of the hottest niches in the AI application field. In the future, it may replace ChatBot as the mainstream application form in multiple fields. Because compared with ChatBot, it can better improve productivity.

ChatBot can help people complete certain parts of the workflow, while Agent can complete the entire workflow and directly deliver the results after receiving instructions. Furthermore, Agent can work in parallel. An experienced professional can collaborate asynchronously with dozens of Agents simultaneously, greatly improving work efficiency.

However, since Agent is a very new application form, there is very little infrastructure specifically built for it. The current mainstream cloud architecture is built for software delivery: stateless services and immutable deployments. This model is very suitable for handling the "outer loop" of production-level code but does not match the actual working state of Agent.

Like humans, Agents need to iterate continuously. They need to install dependencies, write files, run tests, and debug; they also make mistakes, create branches, fix errors, and try repeatedly. They just do all this at machine speed in a large number of parallel threads.

The infrastructure that Agents need requires a new technological paradigm: instant, stateful, and ephemeral temporary computing power. This computing resource cannot just be a disposable sandbox but should also have the feel of a truly programmable computer.

A startup called Daytona has created a new type of "composable computer" for Agents, also known as an "AI sandbox." These sandboxes allow Agents to run code, execute commands, and handle "computer-use" workflows, and they have full control over the underlying environment (including CPU, memory, disk, operating system, etc.). Moreover, the startup and allocation of all these resources can be completed within 60 milliseconds.

Recently, Daytona received a $24 million Series A financing led by FirstMark. Other participants in this round include Pace Capital, existing shareholder Upfront Ventures, E2VC, and strategic investments from Datadog and Figma Ventures. In addition, this round of financing also attracted a group of top angel investors, including Gorkem Yurtseven (co-founder of Fal), Theo Browne (founder of T3 Chat), etc.

Even earlier, Daytona received a $2 million pre-seed financing and a $5 million early-stage financing led by Upfront Ventures. In these two rounds of financing, there were also a large number of angel investors, such as Paul Copplestone (co-founder of Supabase) and Prashanth Chandrasekar (CEO of StackOverflow).

Agents need native infrastructure, not just a simple "sandbox"

Ivan Burazin is the CEO and co-founder of Daytona. He has worked with Daytona's founding team for many years. The team members include Goran Draganić and Vedran Jukic, etc.

As early as 2009, they created the world's first browser-based IDE (PHPanywhere). After that, their careers have also been closely connected with the developer ecosystem. Until 2023, Ivan Burazin and the co-founders jointly founded Daytona.

The founding team members of Daytona, Ivan Burazin, Goran Draganic, and Vedran Jukic. Image source: Daytona

Initially, Daytona was a development environment manager, specifically providing automated development environments for human engineers within large enterprises. Technology giants like Meta and Google already have similar tools internally, but traditional enterprises in the Fortune 500 do not have the in-house technical strength to build such infrastructure.

After the Agent boom emerged, Daytona completely transformed from serving human developers to serving AI Agents.

In essence, Daytona is a "composable computer" tailored for AI Agents. If AI Agents are regarded as digital knowledge workers, then Daytona's products are the laptops or PCs they use to complete their work.

It is called "composable" because the form of this "computer" can be precisely defined at the code level: what type of CPU is needed, how much memory, whether a GPU is required, how much disk space, and even which operating system to run. All these resources can be instantly started within less than 60 milliseconds.

As AI Agent infrastructure, it is very convenient for human programmers to use Daytona. They only need to log in and directly command AI programming tools such as Claude Code or Open Code to complete the configuration.

What are the most core product indicators of AI Agent infrastructure? Daytona believes they are mainly speed and concurrency.

In terms of speed, Daytona has currently achieved an extremely fast cold start of less than 60 milliseconds, making human users hardly feel any delay.

In terms of concurrency, quickly starting a sandbox is one thing, but maintaining an extremely fast startup under large-scale concurrency is another. If users need to conduct reinforcement learning, they may need to start thousands of environments simultaneously in an instant. Daytona has carried out special optimization for this situation.

In terms of the underlying operating mechanism, Daytona has completely self-developed the entire technology stack, which is tailored for AI Agents and runs on Daytona's own physical machine cluster. It neither runs nested within a VM nor uses Kubernetes, Nomad, or any existing open-source orchestration systems. Because these systems were initially designed for large-scale deployment of static applications, not for agent runtimes.

This system's technology covers: strict security boundaries, orchestration systems, resource preheating pools, snapshot mechanisms, resource management, observability, and enterprise-level control.

Currently, there are already many large software platforms in the market that will launch "sandbox" products for Agents, and some teams have also tried to build them internally. However, creating a "dedicated computer" for Agents is not easy. A qualified AI sandbox must have the capabilities of extremely fast startup, full-state maintenance, and long-term operation. Without completely reconstructing the underlying system, it is impossible to meet these strict indicators.

The logic behind this is very simple. Agents are much faster than humans, so their computers must be extremely agile.

To evaluate the capabilities of an AI sandbox, Ivan Burazin, the founder of Daytona, believes that two dimensions need to be considered: primitives (the smallest and most basic building units in a specific system) and toolchains.

In the dimension of primitives: How fast can it start? How much concurrency can it support? Can it run forever? Is it equivalent to a real full-featured machine or virtual machine?

In the dimension of toolchains: Just as a laptop comes with a browser, file manager, and terminal when it leaves the factory, the "computer" of an Agent should also have corresponding native components. Daytona has built-in core tools such as a Git client, LSP (Language Server Protocol), and firewall. These tools can either significantly improve the productivity of Agents or significantly reduce the system operation risks.

In the future, Daytona needs to expand the support scope of operating systems, from Linux and Windows to macOS. On the other hand, it needs to support running in the customer's existing cloud Kubernetes cluster.

The current Daytona is mainly deployed privately locally. In the future, it will support Kubernetes, but it will not use Kubernetes to directly run the sandbox. Instead, it will use Kubernetes to manage the underlying nodes, which is equivalent to building a two-layer precise orchestration system.

Enterprise customers especially need this: they can enjoy strong physical-level isolation and at the same time achieve automatic elastic scaling of computing power.

Daytona's customers mainly include three categories.

The first category is Agentic AI coding companies. These companies have currently received large amounts of financing, and their scenarios have been verified, with a large number of users.

The second category is companies focusing on browser operations or computer operations. Daytona becomes their underlying cornerstone, providing the infrastructure required for their Agents to interact with the human interface.

The third category, and the latest one, is the infrastructure for reinforcement learning (RL) environments. One of the most well-known benchmarks for testing the real capabilities of Agents in the industry, TerminalBench, uses Daytona as its underlying operating framework. Daytona has thus attracted a large number of top teams running RL workloads.

Daytona helps its customers save between 6 and 20 hours in specific workflows (especially in the RL field), achieving a leap in business efficiency.

When Agentic becomes the mainstream AI application form, the growth potential of Agent-native infrastructure is almost infinite

Agentic has become the most obvious trend in the AI application field. Although Chatbot is still the area with the most concentrated user volume in the AI application field, the productivity advantage of Agent over Chatbot is very obvious.

Chatbot can only collaborate with people one-on-one and can only complete one part of people's workflow at a time. Agent, however, can collaborate with people one-to-many, and it can complete the entire workflow end-to-end and directly deliver the results.

Its greater advantage compared with Chatbot lies in its ability to run in parallel, and people can operate it asynchronously. In work organization, the future AI application mode will probably be that a core human employee manages multiple Agents and lets them complete different work tasks in parallel. This form has already been realized in AI-native application companies such as cognition.

We can completely predict that in the future, Agent will become an important part of the mainstream labor force.

Previously, Ivan Zhao, the CEO of Notion, said in a podcast interview that if a software company's product cannot be used by Agents, it will be in danger in the future. Referring to NVIDIA's development trend in the past three years, we can say that "agent compute" will become one of the largest incremental markets in the infrastructure field. Given the characteristic that one person can manage multiple Agents simultaneously, the final scale of this market should exceed the human-centered computing market.

As mentioned before, the infrastructure of mainstream cloud providers is still designed for human computing needs, which has a large gap with the needs of Agents. And although companies like Daytona and large enterprises such as Alibaba and Amazon have seen this direction, their products are still far from mature.

For entrepreneurs, this is obviously good news because on the one hand, the market potential is huge, and on the other hand, there are still few competitors, especially the large enterprises' competitors are relatively slow in innovation speed.

Facing this huge opportunity, we welcome entrepreneurs to join in and strive to become the NVIDIA + AWS of the Agentic era.

This article is from the WeChat official account "Alpha Commune" (ID: alphastartups). The author is the one who discovers extraordinary entrepreneurs. It is published by 36Kr with authorization.