Amazon's AGI Lab unveils Nova Act, a general purpose AI agent to control a web browser and independently perform some simple actions, and a Nova Act SDK
Amazon on Monday unveiled Nova Act, a general-purpose AI agent that can take control of a web browser and independently perform some simple actions.
Context & Ripple Effects
Amazon had already created a dedicated agentic-AI group inside AWS, making Nova Act a concrete product expression of that internal agentic-AI push. It also enters a developing browser-control category in which Google’s Project Mariner prototype demonstrated cursor, clicking and form-filling capabilities.
The accompanying SDK matters because it gives developers a route to build around Amazon’s browser-action capability rather than treating it solely as a standalone demonstration.
First-order effects
- Amazon and developers gain a new SDK-based surface for building browser workflows that can carry out simple actions autonomously.
- Nova Act places Amazon directly alongside browser-operating agents such as Google’s Project Mariner, with the practical scope described here limited to simple tasks.
Second-order effects
- Browser-agent vendors will be judged less on a chat interface and more on whether their agents reliably complete multi-step web workflows; developer tooling becomes a key point of differentiation.
- Teams evaluating repetitive browser work gain another potential implementation path, but the reported simple-action scope means human oversight and narrow workflow selection remain central.
Third-order effects
- If SDKs become the standard delivery mechanism, browser automation may shift from brittle scripted workflows toward application-embedded agents that interpret and act on web interfaces.
- The category’s long-term value will depend on dependable execution across changing websites, making reliability and workflow fit—not merely model capability—the likely competitive boundary.
The trend: AI providers are moving from models that answer questions toward developer platforms that can execute bounded tasks across existing web work surfaces.