Perplexity launched Teammates this week. They're persistent agents that run around the clock, can be assigned roles, work in parallel, and take requests through Slack. Right now they sit on the $200 a month Max plan.
The feature list is fine. What caught our eye is the word "persistent." Most agents people use today live and die inside a single session. You open a chat, ask for something, close the tab. A teammate that stays on and keeps working is a different kind of thing to manage, and most teams haven't built the habits for it yet.
An always-on agent is a hire, not a tool
When an agent only exists for ten minutes, mistakes are cheap and contained. When it runs overnight, in a shared channel, with a role someone typed into a settings box, you've effectively added headcount. Nobody would onboard a new person with "just help out where you can." Yet that's roughly how many teams set up their first agents.
Before turning one on, write it a job description. It doesn't need to be long. Four lines is enough:
- Owner. One named human who gets the blame and the pings. Not "the team."
- Scope. What it can read, what it can write, and which channels it's allowed to speak in. Be boring and specific.
- Budget. A hard monthly cap, in dollars and in runs. An agent working 24/7 has no natural stopping point, so you have to supply one.
- Handoff rule. The exact conditions where it stops and asks a human. Anything involving money, customers, or deleting things is a good starting list.
Parallel agents multiply your review queue
Perplexity says Teammates can run in parallel. That's useful, and it also means the output arrives faster than anyone can read it. We've written before about AI pull requests waiting 4.6x longer for a reviewer. The same thing happens with agent output in Slack. Five agents posting summaries, drafts and status updates is a lot of text for a team of eight.
The fix is the same one that works for code. Decide up front what needs a human look and what doesn't. A weekly competitor digest can go straight to a channel. A customer reply draft cannot. Put the line in writing, and make the agent follow it by design rather than by good behaviour.
Slack is where the permissions get messy
Putting agents in the place where your team already talks is smart for adoption. It's also where context leaks. A private thread about a layoff, a pricing change or an unhappy client is one @mention away from becoming part of an agent's working memory.
Start with a single public channel for the agent. Give it its own account instead of a borrowed human login, so you can see in the audit log exactly what it did. Expand only after a month of watching it behave.
What we'd do this quarter
If you're a founder or a product lead, don't wait for the perfect tool. Pick one repetitive job that has a clear definition of done, like triaging inbound support tags or drafting the Monday metrics summary. Assign one agent, one owner, one budget. Review its output every day for the first two weeks, then weekly.
You'll learn more from that small pilot than from any vendor demo. And when the next always-on agent ships from the next vendor, you'll already have the template: owner, scope, budget, handoff.
We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!