Perplexity launched a set of work-focused AI agents on 5 November 2026. The Verge reported that the product, called Teammates, runs always-on agents that can take on multiple roles inside a workspace, pulling in the projects, skills and memory from the Perplexity tools a user already has. Teammates connects to Perplexity Computer, an AI system that can deploy agents on your device. The Verge framed it as Perplexity taking on OpenAI and Anthropic.

I could not load the full article, so those are the facts I am working from.

The pitch is another agent on the roster

Every few weeks a new agent arrives with a name that sounds like a person. This one takes “roles”. It is always on. It remembers what you did last time. The word that does the work in the announcement is “teammates”, because a teammate is someone you add. Add a researcher. Add an analyst. Add a scheduler. The product is built to make adding feel natural.

I have watched the same pattern with personal agents like Muse and Instinct sitting next to Codex and Claude. Each one is pitched as the agent for a particular job. Put them together and you have a stack. A stack feels like progress, because you can count it.

So I ask the question I always ask first. Is there a need for it? Not what can it do, but what problem does it solve that I could not already solve, and is it worth adding?

What actually changed, and what did not

What changed is packaging. Agents that persist, hold memory, and take on different roles already exist in the tools most operators use. Perplexity now has its own version, wired into its own search and its own device-level system. If you already live in Perplexity, that is a real convenience.

What did not change is the work. A shop still has the same order questions, the same refund rules, the same supplier emails. A clinic still has the same bookings, reminders and no-shows. A marketing team still has the same briefs to write, pages to publish and numbers to check. None of those workflows got simpler because a new agent can be assigned a role. The hard part of any of them is still knowing the business well enough to describe the workflow correctly, and then checking the output.

Memory and roles are features. They are not a reason to switch, and they are not a reason to add.

One tool you know deeply beats five you know a little

I use Claude for almost everything. Research, drafting, building software, counting and checking results. One tool, used well, has been more efficient and more effective for me than any stack I tried before it. The reason is simple. Depth compounds. The more of my work runs through one place, the more context that place holds, the better the briefs get, and the less I have to re-explain.

Spread the same work across five agents and every one of them holds a fifth of the context. You spend your time copying between them and reconciling what each one thinks is true. You end up as the integration layer, which is the job you were trying to get rid of.

Here is the test I would apply to Teammates, or to any new agent. Name one workflow it would run end to end that the tool you already use well cannot. The whole workflow, start to finish, not a step in the middle. If you can name one, and it matters to your business, try it on that workflow only. If you cannot name one, the honest answer is that you have not gone deep enough on the tool you already have.

Most operators I meet are nowhere near the ceiling of the one tool they pay for. They use it to answer questions when it could be running the process. That gap is bigger than any gap between vendors.

Where the workflow runs matters more than who runs it

This is the same view that shapes what I build at AeroChat. A merchant describes the customer service or sales workflow they want in their own words, once, and the system runs it the same way every time, connected to their own order and backend data so it can act, not just answer. The value is a process that runs end to end inside one place. A smarter reply from a new agent does nothing if the refund still has to be keyed in by hand afterwards.

The principle holds whether or not you ever touch my product. The unit of work is the workflow, not the prompt, and not the agent either. Pick the place your workflows will live, then make that place carry as much of the process as it can.

What to do

If you run a shop, a clinic, a service business or a marketing team, do this before you add Teammates or anything like it.

Write down the three workflows that eat the most time each week. For each one, note which steps already run in the AI tool you use and which steps still happen by hand, in a spreadsheet, or in someone’s head.

Take the tool you already have and push those by-hand steps into it. Give it the context it is missing. Fix the brief when the output is wrong. Check what comes out.

Only then look at a new agent, and only against a named workflow the current tool genuinely cannot run. If no such workflow exists, close the tab and go back to the tool you have.

Source: The Verge: Perplexity launches its own take on AI agents for work