Amazon unveils Nova Act, a browser-controlling AI agent

Most AI agents still struggle when you ask them to click a button or fill out a web form. Amazon thinks it has cracked that problem, and it has a built-in Trojan horse ready to deliver it to millions of living rooms.

On Monday, Amazon’s newly formed San Francisco AGI lab introduced Nova Act, a general-purpose AI agent that can take the wheel in a web browser. Alongside the model, Amazon rolled out the Nova Act SDK on nova.amazon.com, letting developers build interactive web automation prototypes.

Under the hood of Amazon's agent push

The project is spearheaded by former OpenAI researchers David Luan, who previously founded Adept, and robotics pioneer Pieter Abbeel, who co-founded Covariant. Amazon poached both last year to lead its frontier agentic push.

Instead of shooting for total autonomy right out of the gate, Nova Act is designed to reliably knock out short, simple tasks like ordering lunch from Sweetgreen, picking dates on a calendar, or submitting forms. Crucially, the SDK allows developers to pinpoint exactly where human review should happen.

Amazon claims Nova Act beats the competition on screen text comprehension. On the ScreenSpot Web Text benchmark, Nova Act scored 94%, outpacing Anthropic’s Claude 3.7 Sonnet at 90% and OpenAI’s Computer Use Agent at 88%. However, Amazon notably skipped broader industry benchmarks like WebVoyager.

For now, the system is strictly a research preview, but its eventual landing spot is much bigger: it will power major features in the long-delayed Alexa+ upgrade.

Why it matters

OpenAI and Anthropic may have built the browser-use category with Operator and Computer Use, but Amazon has something neither of them possesses: hundreds of millions of Alexa-enabled devices already sitting in people's homes.

If Nova Act works reliably, it turns Alexa from an glorified egg timer into an assistant that can actually handle dinner reservations and errands across the web. That distribution gives Amazon an immediate consumer moat that pure-play frontier labs will struggle to match.

We think Amazon's pragmatic focus on narrow, verifiable tasks with built-in human intervention is smart. The biggest hurdle for agents like Claude or Operator has been flaky reliability and bizarre failure modes. By accepting that full autonomy is still out of reach, Amazon is prioritizing day-one usability.

Still, skepticism is warranted. Benchmarking against a selective internal test instead of WebVoyager suggests Nova Act may still face the same edge-case nightmares that plague every browser agent today. The stakes are huge: if Nova Act stumbles when Alexa+ finally launches, it could set consumer trust in autonomous agents back years.