AWS previews the Amazon Nova Act SDK
AWS previewed the Amazon Nova Act SDK for browser automation agents, including IAM, S3 data controls, a browser tool, Python scripting, and testing support.
By Hypler editorial team · Source: AWS.
Original source date: .
Hypler publication: . Hypler updated: .
Estimated reading time: 5 minutes.
Topics: AWS, Nova Act, browser automation, agents

The historical preview focused on browser execution
AWS announced the preview of the Amazon Nova Act SDK on July 16, 2025 for browser automation agents. AWS said it was powered by a Nova Act model fine-tuned for agentic AI and browser execution, with IAM credential management, Amazon S3 data controls, and the Genesis Browser Tool. It positioned the SDK around Python-based, multi-step workflows developers could test and debug before deployment. AWS said the preview was available in US East (N. Virginia) at that time. Preview status and regional launch are historical facts; they do not establish current availability or authority for third-party portal actions.
The browser's state becomes application state
A browser agent works through pages built for people. Its state includes navigation history, cookies, redirects, form fields, downloads, dialogs, and content that can change while it runs. Record task identity, selected account, allowed domains, current step, expected page condition, and terminal outcome. Treat every page as untrusted input, including apparent instructions. The action plan must derive from the approved task contract, with page state validated before entering data or submitting a form. Selectors, assertions, and screenshots can help, while structured target identifiers and approved parameters make the run inspectable.
A read-only portal task is a useful first step
Imagine a hypothetical operation that opens a vendor portal, reads a named order's status, and prepares a draft update for an employee. The task names the order, approved domain, read-only account, and maximum pages to inspect. The agent captures displayed status, retrieval time, and page identifier, then returns a structured draft outside the portal. It stops before any submission. A different account, missing order, changed confirmation dialog, or request for extra credentials becomes a recorded mismatch for the employee. The workflow gains evidence and a draft without creating an external effect.
Consequential steps need checkpoints and receipts
Break the workflow into authenticate, locate target, validate target state, prepare input, request approval, submit, and verify outcome. Each checkpoint needs a success condition and terminal reason. Keep evidence such as URL pattern, target identifier, confirmation text, or receipt identifier while avoiding broad capture of personal or account data. Before a change, create an execution request containing target, parameters, actor, policy result, and idempotency strategy. Verify after submission through a separate observation where possible. A timeout or disconnected session should enter reconciliation, keeping prepared, submitted, and observed-complete outcomes distinct.
A retry can become a duplicate action
A stale button, delayed response, obscured field, or site that changes state before confirming can leave a script unsure what happened. Restarting it can duplicate a submission. Use a safe lookup or receipt check before another attempt. Test login expiry, account switching, changed layouts, slow responses, duplicate submits, unexpected downloads, and adversarial page content. Include a run that loses browser connection immediately after a submit. The correct result is a reconciled state with evidence, rather than an optimistic answer that the action probably succeeded. This is where browser automation differs materially from a tidy API example.
Recovery evidence deserves equal weight
Use controlled accounts or synthetic records where possible. Define normal paths, known exceptions, expired sessions, permission failures, changed page states, and interrupted submissions. Measure completion, intervention rate, false completion rate, duplicate effects, and recovery time. Preserve script version, browser configuration, allowed-domain list, and assertions with each result so later regressions have a meaningful baseline. A test that proves a safe pause can be more valuable than one successful click-through. Browser interfaces change even when the agent code does not, so recovery behavior should be tested deliberately rather than inferred from a happy path.
Keep browser power tied to named work
A first release needs a scoped account, audit trail, kill switch, ambiguity escalation path, and named owner. Review retention for screenshots and browser artifacts, especially when pages contain sensitive information. Expand to new domains, actions, or accounts one approval at a time. Hypler's analysis is that the preview identifies useful control surfaces, while application-specific rules determine whether a browser workflow is acceptable. The announcement does not make third-party actions authorized, predictable, or recoverable. Those properties come from the workflow definition, its evidence, and its ability to stop.
Hypler analysis
Hypler’s view
Browser automation turns cookies, redirects, dialogs, and confirmation pages into production state. A safe design is a named workflow with a known owner and recovery path, not a browser permitted to keep clicking.
Hypler technical analysis
Engineering notes
Browser automation joins model decisions to authenticated third-party interfaces, making browser state and workflow recovery part of the engineering contract. IAM and S3 integration identify useful control surfaces, but a safe browser workflow still needs explicit allowed actions, stable checkpoints, and accountable handling of partial external effects.
What to inspect
- Workflow allowlist
- Name permitted pages, form fields, navigation transitions, download or upload actions, and stop conditions; inspect rejection of an unlisted transition.
- Browser state
- Test expired login, account switch, unexpected modal, changed page layout, and cross-tab navigation; inspect whether the run pauses before ambiguity becomes an action.
- Recovery evidence
- Simulate timeout after a form action and duplicate retry; inspect idempotency handling, captured page state, human handoff, and reconciliation of uncertain outcomes.
Scope and limits. The preview describes an SDK and integrations. It does not make third-party browser actions authorized, predictable, or safely recoverable without application-specific controls.
Terms and definitions
- Execution Evidence
- Records that show whether an approved action was attempted, what inputs and targets it used, and what result occurred.
- Tool Calling
- A model interface pattern where the model requests a structured operation from an external tool.
- Least Privilege
- A security principle that grants only the minimum access needed for a specific task.
- Approval-First Automation
- A workflow that requires review and approval before a consequential state-changing action runs.
- AWS previews Amazon Bedrock AgentCore
AWS previewed Amazon Bedrock AgentCore, modular services for runtime sessions, memory, gateways, browser use, code interpretation, identity, and observability.
- Microsoft previews Deep Research in Azure AI Foundry
Microsoft announced a public preview of Deep Research in Azure AI Foundry Agent Service. The company described an API and SDK offering based on OpenAI's research capability, with Bing Search grounding and source-backed output.
- OpenAI introduces ChatGPT agent
OpenAI announced ChatGPT agent, a product feature that combined web interaction, research, code execution, and document work in a virtual computer environment.

Original source
Historical source briefing with Hypler engineering analysis. Source details reflect the original publication date; implementation examples are hypothetical. Technical analysis expanded October 24, 2026; this does not change the original source event or establish a new source verification. Editorial prose revised October 26, 2026; original event, publication and source-check dates are retained.