RESOURCES / THE EVOLUTION BLOG
Agent SDKs and Frameworks: What They are and How to Spot Them
Alisdair Faulkner
What is an agent SDK?
An agent SDK, or agent framework, is a code library that lets a developer point an AI model at a browser and tell it what to do. The developer writes a goal, and the framework turns it into clicks, typing and page navigation. It can run on the developer's own computer or on hosted browser infrastructure, such as Browserbase or Browser Use Cloud.
These toolkits power a lot of legitimate automation, like testing, data entry and internal workflows. The same tools also let attackers run AI-driven sessions against a site at scale.
How are agent SDKs different from other agent types?
The other three types of agents are consumer products: a person asks for help and an agent acts for them. An agent SDK has no consumer in front of it. A developer decides what it does, and it can run the same task thousands of times.
Agent SDKs also sit closest to traditional bots. Some behave so mechanically that Darwinium classes them as bot driven or as a crawler instead of as an agent.
| IN-BROWSER AGENT | BROWSER EXTENSION AGENT | CLOUD AGENT | AGENT SDK | |
| WHERE THE BROWSER RUNS | The person's computer | The person's computer | The vendor's cloud | A developer's computer or a hosted service |
| NETWORK | Home or office | Home or office | Cloud hosting | Varies |
| WHO CONTROLS THE ACTIONS | The AI assistant the person uses | A separately installed extension | A hosted service | A developer's code |
| HOW A TASK STARTS | The person asks in the browser | The person asks through the extension | The person asks in an app or chat | A script runs |
Which are example agent SDKs that Darwinium detects?
| AGENT SDK | MADE BY |
| Browserbase + Stagehand | Browserbase |
| Stagehand (local) | Browserbase |
| Browser Use Cloud | Browser Use |
| Browser-use (local) | Browser Use |
| Notte | Notte |
| Magnitude | Magnitude |
| Nova Act | Amazon |
An agent SDK that does not match a known name is still classified as automation.
What are the tells of an agent SDK?
Some agent SDKs raise no device warnings at all, so device checks cannot be the whole answer. Their behavior on the page is the most mechanical of any agent type:
- Clicks land in the exact center of buttons, every time.
- Keys are entered instantly, with none of the timing of a person typing.
- The framework adds its own helper code to the page.
- It reacts to page content that people cannot see.
Each sign can be explained away on its own. Together, from different checks, they make a clear case, and that is how Darwinium reaches a decision.
Why does this matter?
An agent SDK has no shopper behind it, only a developer's goal. That goal may be a test suite or a partner integration. It may also be account sign-ups, credential attacks or scraping at volume. Recognizing SDK traffic lets you allow the automation you approve and stop the rest before it reaches your servers.
Frequently Asked Questions
Is an agent SDK the same as a bot?
Not quite. A bot follows a fixed script. An agent SDK uses an AI model to decide each step, so it can adapt to a page it has not seen before. Some agent SDKs still behave so mechanically that they look like ordinary bots.
Can an agent SDK run from a home computer?
Yes. Stagehand and browser-use both run on a developer's own computer as well as on hosted browsers, so the traffic may not come from a data center.
Are agent SDKs always malicious?
No. Many teams use them for testing and approved automation. The point of detection is to know which SDK traffic you have and apply the right policy to it.