Automate legacy web applications with Amazon Bedrock AgentCore Browser Tool
Amazon Web Services Salman Moghal ● Covered by 3 sources
AWS says its new Bedrock Browser Tool lets AI work inside old web apps like a human. That could cut the manual grunt work that still haunts regulated industries.
Based on reporting by Amazon Web Services, Salman Moghal — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Enterprises still living with legacy web apps have a problem that RPA never really solved. The screens look simple enough, but the work behind them is messy: multi-step forms, session quirks, MFA, and systems that expose HTML but no clean API. AWS is pitching Amazon Bedrock AgentCore Browser Tool, paired with Strands Agents, as a way to let AI take over those browser chores without the usual brittle bot drama.
The pitch is aimed squarely at industries like healthcare, manufacturing, retail, and financial services. AWS uses an insurance workflow as the example: tens of thousands of policy changes a year, each one bouncing through several screens to update coverage, beneficiaries, claims, or endorsements. That kind of work eats time and creates errors, which is why it remains a cost center instead of a neat automation win.
The Browser Tool runs a managed Chromium instance in the cloud and gives agents isolated sessions to work in. It connects through Playwright over WebSocket-based Chrome DevTools Protocol, so the agent can drive whatever legacy interface sits on the other end as long as it’s reachable over HTTP or HTTPS. AWS says each session is isolated, then torn down and sanitized when done. Browser profiles can keep sign-in state across sessions, which matters when the target app hides behind SSO or MFA. Proxy settings can also route traffic through corporate infrastructure, including systems behind internal networks or IP-allowlisted portals.
The architecture in the reference implementation is built around that browser layer. A React chat UI loads from Amazon CloudFront and S3. Amazon Cognito handles login, an ALB terminates TLS, NGINX forwards the connection, and Amazon Bedrock AgentCore runtime routes the request to a browser-agent container running Strands Agents and a human-in-the-loop handler. The agent then calls a vision-capable foundation model on Amazon Bedrock, looks at screenshots, decides the next action, and stores transcripts and screenshots in S3. The screenshots are fetched with pre-signed URLs, so the image bytes never travel over the WebSocket.
AWS is also leaning hard on compliance. Session recording, CloudWatch logging, CloudTrail, IAM controls, and audit trails are all part of the pitch, along with the claim that the setup can support regulated industries that need traceability and human oversight. The point is simple: don’t just automate the click path, automate the recordkeeping too. That’s the part most vendors skip until the auditors show up.
My take — AI-written commentary, not fact-checked reporting
This is the kind of automation pitch that actually makes sense: not magic, just a managed browser with memory, logs, and fewer excuses. The real tell is that AWS is targeting the ugly middle of enterprise software, where APIs are missing and people are still paid to click through screens that should’ve retired years ago. Closed, auditable, regulated automation beats another selector-brittleness fairy tale every time.
Read more about this at: Amazon Web Services