Stop Handing Agents a Prompt. Hand Them a Job.

Perplexity just launched Teammates: personal AI agents that own ongoing work instead of answering one prompt and disappearing. They run inside Perplexity Computer, reuse your projects, skills, memory, and connected tools, and can message coworkers in Slack. For now they're available to Max subscribers on the web, with native apps coming later. Engadget and The Verge both put Max at $200 a month.

That price will get most of the attention. The part that matters for operators is the shift in the job description.

What actually changed

Until recently, most business AI was a session. You opened a chat, described a task, got an answer, and started over tomorrow. Teammates are built for the next step: give an agent a responsibility, and it keeps working.

Perplexity's product materials say a teammate can own something like a growth campaign, bug triage, metrics tracking, or market research. It plans the work, brings in more agents when the load grows, and brings finished pieces back for review. Teammates can manage other teammates. They notify you when a decision is needed. Each one gets its own thread so you can check progress, change direction, or take over.

CEO Aravind Srinivas posted that Perplexity already has about 10,000 teammates in its internal workspace from the last 30 days — roughly 20 times its headcount — with about 1,000 active on a recent day. In a follow-up, he said the median employee now gets 7x more agent-hours per week from teammates than in September, and one in five is already at 40 or more agent-hours a week. His longer-term pitch: every org with hundreds of teammates supporting each person.

He also said teammates tend to ask for approval, and that you decide how proactive each one is when you set it up. That detail is more important than the demo videos.

Why it matters for operators

Most small and mid-size businesses are already past "can AI write an email?" The hard part now is whether an AI can own a recurring lane of work without creating a second mess you have to clean up later.

Owning a job is different from answering a prompt. Ownership means the agent watches for changes, follows through across days, and touches the same systems your people use. That only works if three things are clear before day one:

1. What it owns. One lane. Bug triage. Stale customer follow-up. Bid pipeline hygiene. Not "help with sales."

2. What it can touch. Only the tools and records that job needs. Slack for updates is not the same as write access to the CRM and the bank.

3. When a human has to say yes. Outbound messages, money moves, deletions, and anything customer-facing should default to approval until the agent has a track record.

If you skip those three, you don't get leverage. You get a busy intern with your admin password.

My take

At Infraxio, this is the same pattern we use when we put agents into a business hub or a custom system. The agent works from an assigned task, with scoped access, a named owner, and a clear gate before anything leaves the building. That structure is what lets us hand over more work over time, not less.

So if you're evaluating always-on agents this quarter — Perplexity's or anyone else's — don't start with the feature list. Write the job description first. Name the owner. List the systems. Mark the approval points. Then turn the agent on.

The labs will keep shipping agents that can run for days. The operators who get value from them will be the ones who treated each agent like a hire with a role, not a chat window with a credit card.

Sources: Perplexity product page, The Verge, Engadget, @AravSrinivas on X, and follow-up on agent-hours.