Perplexity Computer automation: schedules you keep in an open-source repo

Perplexity Computer keeps working after you close the tab. Its launch post promises workflows "capable of running for hours or even months". The part that repeats is in the help center: Scheduled Tasks, run from the Automations dashboard at perplexity.ai/computer/automations. Both pages were checked in October 2026.

Creating one starts as a conversation. Describe the job and its cadence inside Computer. It proposes a schedule, you confirm, and the task appears on the dashboard with its next run time. A background run then starts a fresh agent with no memory of the conversation that created it. That suits a price check every morning. It does not suit "carry on with the draft I was editing", which Computer routes to an attended run instead.

The documented edges matter before you plan around them. Nothing fires faster than once an hour, so a queue that needs a worker every five minutes sits outside this surface. Each Computer conversation can own fifteen tasks, managed from that conversation or from the dashboard view across all of them. Every run spends credits. A skipped run pauses the task, and the task stays paused once credits return until somebody re-enables it by hand.

The gap here is what the recurring job is made of. On a dashboard the job is a row: a name, a cadence, a next run time, and no author, no diff, no history anyone outside the account can read. Perplexity's own article is candid at the edges. Deleting a task removes it permanently. A paused task will not restart on its own. Each task belongs to the conversation that created it. That is what a schedule looks like when the only copy lives inside one vendor's product, on a plan that vendor prices, and on a runtime that vendor controls.

A committed trigger has a shape you can argue with. The starter's manifest holds the cron expression, the agent, the timezone, the run state and the prompt in one block. The one block it ships uncommented is the harness reflector: cron: "0 0 3 * * *", timezone: UTC, enabled: false, off by default because every run creates a sandbox. The weekday shape sits below it in a commented-out daily-digest example, 0 0 9 * * 1-5 on its own timezone, America/Los_Angeles. Move a block between projects and the job moves with it. Fork the repo and the schedule forks too. When the schema changes, you see it as a diff on your next run rather than a migration nobody announced.

A job fires at nine on Monday whether or not anyone has read its prompt since spring. That is the argument for keeping the prompt in the repository the agent already works in. The edit to the schedule and the edit to the code it touches arrive in the same review, from the same person, at the same time. A run that wants to change something downstream leaves a branch and a proposal instead of a silent edit. The agents page for teams covers who reviews what.

The run itself is a session on its own machine and its own branch. The agent's file is blunt about how long that lasts: "Only what you commit + push survives." That one rule does most of the operational work. A monitor that finds nothing new leaves nothing behind. A monitor that finds something leaves a branch somebody can read. No dashboard history to reconstruct, no notification to take on trust.

Owning the schedule means owning the boring parts. You keep the manifest valid as the schema moves. You keep secret values in Secrets Manager while their names sit in the repo. You keep an image the sandbox can boot, and you answer the review queue your own triggers fill. Nobody patches your kortix.yaml for you, and nothing merges itself.

For work that matters, that is a fair trade. For work that does not, it is a poor one. A weekly report someone skims is fine in a dashboard. A weekday job that reads production data, writes into a customer's file or opens a change request against live code deserves a file, a diff and a named owner. Kortix is open source, so the runtime that fires the trigger and the manifest that declares it are both readable, and the self-hosting guide covers running it on a server you keep.