Automation and Agents
How to Build Recurring Web Automations with Perplexity Computer
A proposed weekly pricing workflow that makes changed terms, missing checks and review effort visible.

On this page
A weekly price report can arrive on time and still compare the wrong numbers. An annual price expressed per month, a monthly subscription and a temporary discount can look interchangeable in a summary. Preserve their terms and the account lead can identify a real change. Lose those terms and a polished report can send the briefing in the wrong direction.
Perplexity's September 29, 2026 Automations announcement describes standing assignments that run on schedules or specified events in connected apps and build on previous work. You can create one through Computer's omnibar or Automations, name its sources and destination, and specify review points. That gives you a way to delegate recurring collection and comparison. The quality of the result still needs an acceptance test.
This guide follows a proposed weekly pricing report for a small agency. The prices and report examples are hypothetical. We haven't run this assignment or measured Computer's performance. The goal is a report whose sources, units, changes and incomplete checks are clear enough for an account lead to inspect. The same design can help with other recurring web research, provided you replace the example's comparison rules with rules appropriate to that job.
Start by writing the decision your report should support. For the proposed agency, it is whether a competitor changed a public plan price or an included limit enough to warrant updating a client briefing. It is not whether the agency should match that price, change its own offer or contact a customer. Those decisions require information beyond a competitor's page.
Choose a small set of relevant pricing pages and name the plans you care about. Record why each one belongs in the report. A direct competitor's plan may matter more than a larger company's unrelated enterprise offering. When someone suggests adding a source, ask which decision it improves. A larger source list creates more opportunities for missing data and mismatched comparisons, even when collecting the pages feels easy.
View image detailFor each plan, define the fields needed to understand its price: amount, currency, billing period, commitment period, seat basis, important included limits and whether the displayed offer is promotional. A price labeled per user is different from a price for the whole account. A headline monthly equivalent may require an annual payment. The report should preserve the source's terms before anyone tries to compare costs.
Define material change in the agency's own language. A changed seat limit might matter even if the headline price stays the same. A new illustration or reordered feature list might not. Include ambiguous changes in a separate question for the reviewer instead of asking the agent to silently decide their significance. This keeps the report focused without pretending that every business consequence can be inferred from the page.
Name the reader and the destination at this stage. A report for the account lead should lead with changes that could affect a briefing, then provide evidence. A report for finance might prioritize commitment periods and minimum spend. Those are different assignments. The destination should be somewhere the named reader will actually review, with access appropriate to the information it contains.
A weekly schedule fits this pricing example because the agency makes its comparison during a weekly review. Choose a time that leaves room to inspect the report before that meeting. Record the intended timezone and verify the schedule shown in your workspace. Avoid copying a cron expression from an unofficial tutorial unless the current product actually documents that control.
An event is useful when work should begin because something happened in a connected app. Perplexity names Slack, Gmail, Outlook, Linear and GitHub in its launch description and says conditions can narrow the events that start an assignment. A new customer request and a Monday pricing review are different signals. Don't introduce an email trigger merely because it looks more autonomous than a schedule.
View image detailConsider the decision delay you can tolerate. If a pricing report only informs Friday's meeting, an hourly schedule may produce many reports no one uses. If a change requires immediate investigation, first decide which source supplies a suitable event and whether that event is supported in your workspace. A public page changing is not automatically a documented Computer webhook event.
For historical context, RuntimeWire's September 22 investigation examined Gmail-event code in a desktop package behind an internal allowlist. It preceded the launch. That reporting helps explain the move from periodic checking to event-triggered work, but its internal route names are not a public Automations API. Use the current product description and your available controls for setup.
Decide what a quiet week should produce. The lead needs a completed check with no material change, or an explicit incomplete result. If no report arrives, inspect status and history before interpreting the silence. Name the person responsible for that check so an unattended destination doesn't hide a missed run.
Perplexity's documented entry points are the omnibar and Automations followed by New Automation. Describe the job, choose the schedule or event, connect the relevant accounts and Projects, and identify the destination. The launch also says users can specify which actions may happen autonomously and which require review. Check the settings your workspace exposes rather than assuming every connector offers the same scope.
For this pilot, the following is a proposed instruction to enter as ordinary task text. It is not API configuration: “Every Monday, inspect the named public pricing pages in this Project. Compare the selected plans with the most recent report the account lead confirmed. Preserve currency, billing term, commitment, seat basis and included limits. Lead with material changes, attach direct evidence links, and separate unavailable pages or ambiguous terms. Prepare the report for the account lead's review. Do not change our pricing or contact customers.”
View image detailThat instruction works backward from the finished result. The lead can inspect a change, find its source and see what remains uncertain. The source list should be maintained in the connected Project or another supported location that the assignment can access. Explain which record is authoritative if several documents contain similar plan names. “Use our latest notes” is too vague when old research and current instructions coexist.
Decide what routine output is allowed. A private report or draft can be a sensible first destination. An internal Slack post may also be appropriate if the team has authorized that channel and its audience. State the actual permitted action, rather than labeling every output harmless. A report built from public prices could still include confidential client notes if the assignment reads a broader Project.
Keep the reporting scope independent from the action scope. The agent may discover an unexpected price without gaining permission to purchase the plan, update a client contract or send a recommendation externally. The companion governance guide develops those authority decisions. Here, the pilot's endpoint remains a checked internal comparison.
The first report is an inventory. It can show the current public offers, but it cannot truthfully report a change from a baseline that doesn't exist. Ask the account lead to review the inventory and note which observations are confirmed, which need clarification and which sources couldn't be read. Then identify that reviewed version as the comparison reference for the next run.
Suppose a hypothetical plan displays “$24 per user per month, billed annually.” Preserve that whole statement. A later page displaying “$29 monthly” might describe a different payment option rather than an increase in the annual offer. The reviewer needs the billing and commitment terms for both observations. If the report cannot establish that they are comparable, it should identify the changed presentation and leave the price-change question unresolved. Calculating a percentage first would give the ambiguity a misleading appearance of precision.
View image detailGive each observation a source link and capture time. A current link lets a reviewer open the evidence; the time explains when the report saw it. If a consequential change must remain explainable after the page changes again, retain an appropriate source record in the organization's approved system. Check your rights and data policy before saving material, especially when a source contains account-specific information.
Perplexity says Automations use prior work. That continuity can carry the previous report into a new comparison, but the announcement doesn't specify a user-facing database schema or retention guarantee. Keep the confirmed baseline recognizable in your operating record. Ask the report to name the reference it used. If the reference is missing, produce a new inventory or request clarification rather than inventing a previous value.
Correct a bad baseline explicitly. If the lead finds that the earlier report confused a promotional rate with a standard price, record the correction and tell the assignment which version to use. Preserve the reason for the change so a future reader understands why the comparison shifted. An agent remembering earlier work is useful only when the team can distinguish accepted findings from errors awaiting correction.
A useful change report starts with the decision-bearing finding, then shows its basis. For a hypothetical limit change, it might say the selected plan now advertises fewer included projects, identify the earlier confirmed limit and link the current page. The lead can decide whether that difference belongs in a briefing. The report shouldn't turn it into an unsupported claim about the competitor's strategy or customer experience.
Use three distinct outcomes: checked with no material change, changed with supporting evidence, and incomplete or ambiguous. A failed page belongs in the last category. If one target is unavailable while the other targets look unchanged, the whole result should not read as an unqualified “nothing changed.” Lead with the missing coverage when it affects the reader's decision.
View image detailWhen sources disagree, preserve the disagreement. A public pricing page and an older help article might show different limits. Prefer the source specified for the assignment, but flag the conflicting statement rather than quietly averaging the numbers. If the official page says to contact sales, record that condition. A missing public price is not an invitation to estimate a quote.
Choose a table when the same fields recur across several plans. Include the item, previous confirmed terms, current terms, source and observation time. Explain an important discrepancy in prose beside the table. Avoid making the lead decode a large spreadsheet just to discover that two rows have different currencies. The report's layout should shorten the review while preserving the evidence needed to challenge it.
Keep a separate list of sources checked and sources skipped. This exposes coverage without burying the finding. If the assignment follows a replacement link, ask it to explain why that page is the appropriate successor. A renamed plan may be a new offer, a renamed old offer or an unclear relationship. The report should identify that question rather than treating matching page positions as proof of identity.
Test the assignment with a small set of inputs before increasing its scope. The proposed pilot has a baseline case, a known-change case and an exception case. These are test plans, not results we observed. Keep the instructions, sources and output expectations fixed for each case so you can identify what a revision changed.
For the baseline case, review every selected source and compare the report with the actual page. Check the price terms, plan identity, important limits, evidence links and destination. Record missing or invented fields. A report that looks polished but omits commitment terms hasn't met the specification. Correct the relevant instruction before relying on it for later comparisons.
View image detailFor the known-change case, use a harmless source you control with a deliberate change, or inspect a genuine change in a permitted public source. The reviewer needs to know the expected difference independently. Check whether the report identifies that difference, preserves unchanged fields and uses the correct baseline. One correct case establishes that case's result, not general reliability across every site.
For the exception case, provide a controlled unavailable source or ambiguous billing statement. The expected result is an explicit gap or question without a fabricated value. Don't disrupt a shared production connection to manufacture a test. If you cannot safely produce the exception, record it as untested and choose another source you can control.
Perplexity says run history shows triggers and produced results. Use it to compare what started the work with what arrived at the destination. If a scheduled run is absent, inspect status and the intended schedule. If a run occurred but the report is weak, investigate instructions, access and evidence. Those are different failures. Keep the pilot record short enough that another teammate can repeat the checks.
The October 5 changelog confirms recurring and event-driven web work, previous-run context and review before consequential actions. The launch says waiting for a trigger consumes no credits, while executing an assignment does. That distinction doesn't supply a per-run price or a savings estimate. Examine actual workspace usage before choosing a recurring cadence.
Measure the work the team cares about during the pilot: complete source checks, correctly identified material changes, unsupported findings, incomplete checks and the time a reviewer spends fixing or verifying the report. Record the number of opportunities as well as the successes. Three correct changes among three available changes means something different from three correct changes among twenty.
View image detailInclude maintenance effort in the comparison with your existing process. Someone must update target pages, revise plan definitions and review failures. A recurring assignment can shift effort from collection toward verification without eliminating it. Decide whether that shift helps the account lead make a better briefing. The answer may differ for a stable public pricing page and a frequently changing authenticated portal.
If the workflow is too noisy, identify why. A broad materiality rule, inconsistent baseline or unnecessary schedule can each create extra review. Narrow the task that causes the problem rather than adding a longer instruction full of unrelated restrictions. If the report repeatedly fails to read a source, consider a more suitable official source or a separate documented data integration.
For the pilot, keep a simple review log with the report reference, the disputed field, the correction and the time needed to reach a decision. Separate routine verification from repair. Opening a cited page to confirm a material change is expected review work; reconstructing missing units or hunting for an omitted source is a defect in the report. That distinction helps you decide whether another instruction would improve the assignment or merely move more work onto the reviewer.
An agent is also not automatically the best collector for every job. A stable, authorized data feed may give you clearer structured inputs. Computer's standing assignment is attractive when the task benefits from reading several sources and preparing a contextual result, but that is a design judgment to test. Don't replace an effective deterministic feed solely because a recurring agent can produce a similar-looking report.
Assign ownership of the source list, baseline and report destination. Review them when a competitor renames a plan, a client briefing changes purpose or the account lead leaves. The same recurring instructions can become inappropriate even while the automation continues to run. A job that no longer informs a decision should be paused or retired.
Perplexity says users can edit or pause Automations. Use those controls when the task needs repair, then verify the corrected assignment on the relevant case before expanding again. Editing an instruction doesn't establish that the next run will interpret it as intended. Keep the previous version and the reason for the change in your operating record.
View image detailFor the agency, a handoff should include the selected pages and plans, accepted baseline, intended schedule, destination, review boundary and unresolved source questions. Rise's Second Brain template can organize reference notes and tasks for that handoff. It doesn't supply Computer integration, secure credentials or a guaranteed audit system. Use your organization's appropriate storage for operational records.
Retire connected access that no longer serves the assignment. An old task can retain a misleading baseline or an obsolete destination. Confirm that the owner knows what was paused and which reports should no longer be treated as current. If another person creates a replacement assignment, carry over the useful specification and reviewed evidence rather than an unexplained archive of generated summaries.
Start with one weekly report and a finite pilot. The deliverable to aim for is a comparison the account lead can accept, question or reject with direct evidence. Once the source terms, baseline and exceptions are clear, the team has a basis for deciding whether to increase scope. The useful outcome is a better recurring decision, supported by a report whose limits remain visible.
Checked for this article



