In recent months, at least four AI companies have shipped agents that work in the background instead of waiting for a prompt. Anthropic previewed Claude Duet in October, OpenAI has Dots, and on November 5 Perplexity added “Teammates,” which The Verge describes as always-on agents that take different roles in a workspace. The pattern is clear enough to analyze, though the evidence for how well these products work is thin. This piece sets out what we know, what remains unproven and what to watch.

What the products have in common

Based on reporting from The Verge and The Decoder, the products share four traits:

  • They run continuously. Teammates keep working around the clock, per The Verge, and Duet works on its own cloud computer so a user’s devices need not stay on.
  • They hold memory and context. Perplexity says Teammates pull existing projects, skills and memory from tools a user already has; Anthropic says Duet carries over memory, skills and connections from a Claude account.
  • They connect to work apps. Perplexity’s agents can be tracked in the app and used in Slack. Duet reaches Gmail, Slack and Stripe through connectors.
  • They keep a human approval step for sensitive acts. Duet drafts and prepares, then asks before sending, paying or deleting, our earlier coverage found.

Where they differ

Price and scope vary. The Verge reports that Perplexity’s Teammates require its $200-a-month Max plan, which it calls much more expensive than comparable offers. Duet began as a US-only waitlist for Pro and Max subscribers. Each Teammate gets its own thread, and Perplexity’s chief executive, Aravind Srinivas, said on X that over time he sees “hundreds of teammates working for each person.” That is a vision statement, not a measured result. See our full report on Claude Duet for its approval design.

A hand holding a black smartphone
Most of these agents are reached from a phone app or chat thread; the photo is illustrative. Photo: Dennis Cortés cortes / Wikimedia Commons, CC0

What is still unproven

We have not seen independent tests of how often these agents finish work correctly, how often they need correction or what a typical user spends. Three gaps stand out:

  1. Reliability. Multi-step tasks fail in small ways that compound, and vendors rarely publish error rates. Our guide to evaluating AI agents lists metrics to ask for.
  2. Security. An agent with standing access to email and payments is a target for prompt injection. Tools such as 1Password’s agent integration try to keep credentials hidden, but they are new.
  3. Access to the web. Some sites are blocking agents, as our report on Yelp, Delta and banks shows, which limits what an agent can do on a user’s behalf.

A trend, not a forecast

Two reasonable readings are possible. One is that always-on agents become a standard feature of paid AI plans, like web search did. The other is that they stay niche because trust, cost and reliability limit them to narrow jobs. Evidence so far, such as Anthropic’s read-only Personal Finance feature, points to cautious scoping: vendors grant limited powers first. We do not predict which reading wins; we suggest watching the signals below.

How to evaluate one before you pay

  • Start with a low-risk task, such as summarizing messages, and review every output.
  • Ask what the agent can do without approval, and whether you can revoke access quickly.
  • Check what happens to your data and whether it is used for training.
  • Compare monthly price with the hours it saves in your own workflow.

For fundamentals, read our explainer on what AI agents are.

What to watch next

  • Independent evaluations or user studies measuring task completion and errors.
  • Whether prices fall below the $200 tier Perplexity uses.
  • Security incidents involving agents with standing access.
  • Rules from sites and payment providers on agent access.
  • Whether more vendors add approval and audit features.

Key takeaways

  • Claude Duet, OpenAI Dots and Perplexity Teammates all aim at background work.
  • Perplexity’s version requires a $200-a-month plan, per The Verge.
  • No independent data yet shows how reliable they are.
  • Security and site blocking are the practical constraints.
  • Treat vendor claims about “hundreds of teammates” as vision, not evidence.

Frequently asked questions

What is an always-on AI agent?

It is an assistant that keeps working on tasks in the background, often with memory and app connections, instead of answering only when you type a prompt.

How much does Perplexity Teammates cost?

The Verge reports it requires Perplexity’s $200-a-month Max plan.

Are always-on agents safe to give access to my email?

They add risk because they hold standing access. Prefer products that ask approval for sending, paying or deleting, and start with limited permissions.