Perplexity CEO Aravind Srinivas: The Company is Like a Car, Soon to be Autonomous with the Right Automation
Perplexity CEO Aravind Srinivas stated that the company is like a car, and with the right automation, it will soon operate like a self-driving vehicle; currently, this can be achieved with Perplexity Computer, but identifying the correct automation still requires a considerable degree of human skill and initiative.
The company simultaneously launched Automations within Computer: designed for ongoing work, actions can be triggered by events or scheduled, and can invoke memory, skills, and connected applications like Slack, Gmail, Outlook, and Linear. The help center places scheduled tasks on the automations panel, allowing the cloud to run even when the notebook is closed, writing results into task sessions and notifying users. Users describe tasks in natural language, and the system proposes schedules for confirmation, such as summarizing emails at 8 AM on workdays, checking investor replies hourly, or sending the Salesforce pipeline to Slack on Mondays.
Computer will be launched in February 2026 as a multi-model digital worker, claiming to break down goals into parallel sub-agents, with tasks using a real file system, browser, and tools in an isolated environment, capable of running for hours or even months. The orchestration layer was previously written to call about 19 to 20 models, distributing reasoning, retrieval, images, and videos among different model families. Max subscribers will be the first to use it; there is also a Personal Computer that allows orchestration to be applied to local files and applications. The product page lists pre-made rhythms such as email summaries, competitor comparisons, dashboards, and project tracking.
The official blog states that automation combined with project memory allows subsequent operations to continue from the last context, rather than treating each as a new task. The security side recently announced a red team test of the sandbox SPACE that hosts Computer: granting root access to various model virtual machines and even source code, fixing discovered issues and comparing with other sandbox providers.
In market mechanisms, the purchase is aimed at removing repetitive labor from calendar meetings for knowledge teams, while selling cloud workers that operate continuously based on schedules and events. Funding flows to Max and enterprise seats; beneficiaries are the orchestration layer that has connected to email and tickets, while the pressured side is the chat box and traditional iPaaS connection tools that can only wait for users to ask again. Event-driven functionality is defined by the founder, not new financing.
The self-driving metaphor shifts human work from driving to route selection: agents can accelerate, deciding which automation is worth setting up, and who is responsible if a detour occurs, still falls on the human side. Skills and initiative are left as scarce inputs, rather than being eliminated by the product.
Source: Public Information
ABAB AI Insight
Srinivas has transformed the search box into Computer, and then made Computer into a worker that operates on a schedule; the path is the same: answers are not enough, it needs to do the work for you. Multi-model orchestration is his differentiation—he does not rely on a single lab, treating competitor models as parts of an orchestra. The red team post for sandbox SPACE indicates they understand that once a real browser and real file system are triggered on a schedule, security incidents will recur. Historically, Zapier sold triggers to business personnel, RPA sold keystrokes to finance; Computer connects triggers to agents capable of writing documents and opening web pages, with higher unit prices and permissions.
The capital path is the backend duration in subscriptions. One-time conversations are charged by the number of interactions, while automation is billed daily or hourly in the cloud, with the Max tier justifying continuous operation as a package deal. Connecting Gmail and Linear is to ensure triggers have real events, not just alarms. The motivation is that the growth of search Q&A will encounter "ask and leave," and continuous work can lock users into projects and memory. The enterprise version sells orchestration to teams wanting to reduce coordination meetings, competing for the same "less meeting" budget with Deel-style internal agents and Cursor-style window agents.
The analogy is Unix cron combined with an intern who can read emails, and Tesla describing assisted driving as the stage before full autonomy. The industry phase is agents moving from conversation to duty: whoever stabilizes event triggers first will dominate the calendar. If OpenAI and Anthropic only wait for people to give commands in conversations, they will leave "the 8 AM task already done" to the orchestration layer. Srinivas retains human skills, fearing clients might write off failures as product defects; he also acknowledges that choosing the right automation is closer to management work than simply writing prompts.
Structural judgment is that technology replaces. The mechanism is that once repetitive work can be described by schedules and events, coordination meetings and manual refreshes become removable middle layers. Replacement occurs, not because models suddenly manage companies, but because triggers change agents from "waiting for you to ask" to "doing it on time." The next sentence about the company being like a car is: the steering wheel is still there, but most stretches of road no longer require both hands.