AWS shows how Amazon Nova Act can automate competitor price monitoring

Competitor price monitoring no longer has to be done by hand: AWS has shown how to build an agent on Amazon Nova Act that opens retailer sites on its own, finds the right product, extracts the price and promotion terms, and consolidates everything into a structured result. The idea is simple: instead of employees spending hours switching between tabs and spreadsheets, a company gets a near-continuous view of the market and can make pricing decisions on fresh data.

The problem with manual price monitoring

AWS describes a typical problem for e-commerce teams: to understand how competitors behave, you have to regularly check dozens of product pages, record prices, discounts and promotion deadlines by hand, and then copy them into spreadsheets. The process is slow, scales poorly, and inevitably produces data-entry errors. If a market price changes several times a day, even a delay of a few hours turns analytics into an archive rather than a tool for quick decisions. AWS also notes that the same pain is not limited to online stores: insurers, banks, and travel and hospitality companies also need to keep comparing competitors' offers.

What is Amazon Nova Act

The key element of the solution is Amazon Nova Act, an open-source SDK for browser automation driven by natural-language instructions. A developer builds a Python workflow from small actions: open a site, find a product, go to its page, pull out the needed fields, check conditions, handle an error, or pause. Unlike rigid scripts based on CSS selectors, this approach is designed for live sites, where banners, promo blocks, element order and navigation change constantly. For data extraction, AWS suggests using act_get() with a Pydantic schema, so the agent returns an already validated, typed structure that can be loaded into internal systems, dashboards or pricing models.

Scaling and parallel execution

The example puts particular emphasis on scale. One Nova Act instance works with one browser, but several instances can be run in parallel. In the AWS example, ThreadPoolExecutor distributes the checks across several sources at once.

In practice, this means the agent can sweep Amazon, Target, Best Buy, Costco or any other set of sites in a single pass and then merge the results into one table. In the demo scenario, the user enters a product name and SKU, after which the agent finds the relevant product page, tells sponsored results from organic ones, and extracts the price, promotion details, availability and additional metadata. The output is written to CSV so it can be fed into a BI system, an internal API or dynamic-pricing logic.

Practical limitations

AWS does not gloss over the practical limitations. If a site shows a CAPTCHA, Nova Act does not try to solve it automatically. Instead, the workflow can detect the CAPTCHA and stop so that a person can complete the check manually. For local runs, headed mode is suggested; in the cloud scenario, human-in-the-loop through AgentCore Browser Tool, with interface takeover in the AWS console. The service also returns errors as ActError, so you can build retries, fallback branches and proper logging. For development, AWS recommends extensions for Kiro, VS Code and Cursor, and for monitoring runs, the Nova Act console with traces, screenshots, logs and artifacts in Amazon S3.

What this means

AWS is pitching Nova Act not as an "agent for the sake of an agent" demo, but as a practical layer for routine web processes where speed, reproducibility and scale matter. For retail, it is a direct path from manual monitoring to near-continuous competitive price intelligence. For other industries the signal is the same: if competitor data is still collected through tabs and Excel, browser agents are starting to look less like an experiment and more like a working tool.

In brief:

• AWS showed how to build an agent on Amazon Nova Act for automatic competitor price monitoring.
• Nova Act is an open-source SDK for browser automation; data is extracted through act_get() with a Pydantic schema, and runs execute in parallel and output to CSV.
• When it hits a CAPTCHA, the agent stops for a manual check; errors are returned as ActError for retries and logging.