Opening three browser windows is easy. Knowing which task each window owns, whether a record was already processed and where to resume after a failure is the engineering work.
Multiple windows can help a workflow separate activities, but they can also repeat work or mix account state. Decide first why concurrency is useful. Some steps can run independently. Others must wait for a confirmed result from an earlier step.
Give every task one owner
Imagine a fictional research operation that checks permitted business records and prepares a review sheet. A coordinator could assign independent tasks to three workers. Each task needs an identity and a clear state such as waiting, running, completed or requiring review.
Do not use the fact that a window is open as proof that it owns a task. If the application restarts, the coordinator must determine what was completed before assigning work again. Keep ownership in the workflow's records, with a deliberate recovery procedure.
Keep sessions and destinations distinct
Define which account and system each worker may use. Check the destination before reading or submitting data, and stop if authentication or page structure changes unexpectedly. Do not automate around access restrictions or assume that a visible page grants permission for bulk collection.
For the fictional example, a worker might collect an approved record while another validates an earlier result. That division is a design illustration, not a statement about how HINTX's MLS project assigns its three owner-reported Chrome windows.
Confirm results before progressing
After an export or update, verify the expected output and record its relationship to the task. A click succeeding is different from the business operation succeeding. A downloaded file should be checked before a later step treats it as current data.
Use bounded retries and an explicit unresolved state. If a browser fails after submission, do not assume the operation failed and repeat it blindly. Resume from a verified checkpoint or ask for review when the outcome cannot be established.
Keep AI at a defined boundary
A model may help interpret unstructured material or prepare a summary. It should not quietly change the permitted data source, invent missing fields or turn an uncertain observation into a completed task. The distinction between a fixed workflow and an agent choosing its own steps helps make that boundary visible. Background.
Require review before consequential actions where the business has not authorised automatic execution. Tool approval mechanisms exist, but they need a named reviewer and a specific proposed action. Example.
Measure the complete run
Compare useful completed records, correction work and recovery effort. More windows may increase load or make errors harder to trace. Test a dropped connection, an expired session, a duplicate task and a partial export before expanding concurrency.
Our MLS portfolio distinguishes genuine screens, illustrative workflows and owner-reported operating details. The n8n guide covers orchestration in a different execution environment.
Separate windows from isolated sessions
Three tabs in one signed-in browser may share cookies and account state. Separate windows are therefore not, by themselves, an isolation strategy. Choose the isolation boundary deliberately and document which accounts, downloads and temporary files belong to each worker.
For example, Playwright browser contexts provide separate session environments. That is a useful implementation option, not a claim that a particular HINTX project uses Playwright or that contexts solve every coordination problem. The workflow still needs its own task ownership and recovery rules.
A three-worker design example
Consider an authorised research workflow with independent records to check. The coordinator assigns different record IDs to three workers. Each worker reads its allowed source, validates the result and prepares a review item. A separate dependent export waits until its required records have reached a verified state. This is an illustrative architecture; it is not a reconstruction of private MLS operating logic.
| Stage | Evidence to retain | Recovery rule |
|---|---|---|
| Claim work | Task ID, current owner and start time | Resolve abandoned ownership before reassignment |
| Read the source | Allowed destination, record ID and observation time | Stop for changed authentication or unexpected page structure |
| Validate | Required fields and explicit missing information | Send an incomplete result to review rather than inventing values |
| Save or export | Output reference and a verified completion result | Check for an existing output before retrying an uncertain operation |
| Resume | Last verified checkpoint and outstanding action | Continue only when the prior operation's outcome is known |
Test the business result as well as the click
Browser tools can check whether an element is ready for interaction. Playwright's actionability checks, for example, cover conditions such as visibility and whether a button can receive input. Those checks do not establish that the receiving business system saved the intended record. Verify the resulting record or supported confirmation separately.
Begin with one worker and a small representative sample, then compare the same workload with more workers. Record accepted outputs, duplicates, incomplete results, correction effort and time spent recovering. If the source restricts concurrency, respect that limit rather than adding windows to bypass it. Use an official API or supported export where it meets the need more reliably.
A useful handover includes the permitted sources, account boundaries, stop conditions, output location and the operator who resolves uncertain tasks. Those details make the workflow supportable after the demonstration and give a buyer concrete questions to ask before commissioning automation.
Common questions
Are three windows always faster than one? No. Dependencies, provider limits and verification work determine whether parallelism helps.
Can a workflow recover without repeating everything? It can be designed around verified checkpoints, but that behaviour must be tested with the actual operations.
Show HINTX the workflow you want to automate, including its access permissions and failure cases. We can help decide which steps should run together and which should remain sequential.
Cover image: original AI-generated editorial illustration. It does not depict a client system or measured business results.
Login to comment
To post a comment, you must be logged in. Please login. Login
Comments (0)