Not a Task but a Job: What Perplexity Computer Automations Change

Perplexity added a capability called automations to its Computer tool on September 29, built around a single sentence: the AI is now given an ongoing job , not a one-shot task . Under the old setup the user asked for something, the system did it, stopped, and waited for the next command. Under the new setup the job is defined once and the system keeps repeating it on its own for weeks. According to Perplexity's official announcement, every run inherits context left behind in connected files and earlier work.

The shortest way to grasp the change is a two-week thought experiment. In week one the system collects competitor prices into a spreadsheet. An old-style agent would run the same search from zero in week two, leaving the user to place the two tables side by side. The new automation instead opens week one's file on the second round, compares the figures, and summarizes only the items that changed . In the narrator's analogy, it resembles a worker who takes notes every day and reads them back before starting each morning.

Task schedulers are nothing new, and rivals offer similar timers, so the honest question is whether the memory claim holds up in independent testing. What the launch materials describe is closer to continuity than to recall: runs live inside a project that keeps conversations, files, and connector outputs together. GadgetBond reports that a shared file system plus the Brain memory layer lets later runs start from accumulated context instead of a blank page. Hourly, daily, or weekly cycles can then publish summaries or refresh a status document in the background.

The coverage spans work apps people already use: Gmail, Outlook, Slack, Linear, and GitHub. The system watches for movement in these sources and drops the output where it belongs, such as a message in a Slack channel or a ready draft inside Gmail. Perplexity's announcement post illustrates the pattern with a decision request from a specific customer, summarized against earlier correspondence and the account plan, then forwarded as a notification. Which sources to read and where results go are chosen when the job is defined.

Setup runs through Computer's omnibar or the new-automation screen, following Perplexity's usage guide. A schedule suits regular assignments such as a Monday morning briefing, while an event connector starts work when something happens, like a new message landing in the inbox. Conditions keep the scope tight, for instance limiting review to mail from one particular customer. Instructions, status, run history, and outputs share one screen, and an automation can be edited, paused, or fired immediately as priorities shift.

The first recipe targets incoming leads. When a message about membership arrives, the system checks what this person has already been told, prepares a friendly reply explaining the contents and the time savings, and asks for approval before sending. The result is a finished draft waiting in the inbox for every new enquiry. The critical detail is that sending authority stays with the user.

The second and third recipes are weekly monitoring jobs. One scans the community's past week every Monday: the most asked topics, the tools in use, and the shift since the previous week, condensed into a short brief with items needing a personal reply flagged. The other looks outward, scanning large AI communities and creators for the topics drawing the most attention and suggesting a single content idea for the community. Both rest on the same principle: report the difference instead of re-scanning from zero each week.

The fourth recipe writes member-winning copy: a landing page describing coaching calls, daily tutorials, roadmaps, and a community of business owners in plain language. The fifth recipe catches things that were supposed to happen and did not: a newcomer who joined but posted no introduction and asked nothing within three days is flagged so the owner can reach out personally. Nobody slips through the cracks. The narrator presents this last example as the clearest proof of what makes the update different.

There is an honest caveat in the launch: the system does not run wild. Research, checking, comparing, and drafting may proceed automatically, yet steps such as sending, publishing, or making real changes can require user confirmation. The picture is closer to an assistant who prepares everything but still asks before pressing the critical button. Metering is equally clear: sitting and waiting for a trigger costs nothing, and the counter moves only when the job actually runs. Perplexity's help pages confirm automations can be watched, paused, and re-run from the same screen.

The calendar side shows September as a three-step sequence. On September 17 Perplexity introduced four effort levels from Light to Ultra; an orchestrator plans the assignment and fans pieces out to sub-agents on different models. On September 24 portable Computer opened to AMD Ryzen AI Max chips and the Halo developer platform, where local inference spends no credits and Qwen 3.8 27B plus PPLX 27B install in one click. AMD's newsroom notes the local model keeps routine file work on device while escalating to the cloud for deeper research when needed. The September 29 automations close the arc: the chatbot answers, the agent performs tasks, and the automation holds the job.