Perplexity CEO Aravind Srinivas: About 10,000 AI teammates are already in the internal workspace, 20 times the number of human employees
On November 5, Perplexity launched Teammates, persistent AI agents based on the Computer system that can be assigned roles, goals, and permissions and then autonomously take on ongoing tasks (deep research, data analysis, QA, ticket routing, code review, etc.), while collaborating with other agents and human colleagues through Slack. In the original post, CEO Aravind Srinivas said that nearly 10,000 Teammates had been created in the company’s internal workspace over the past 30 days, reaching 20 times the scale of its human employees, and that about 1,000 were active yesterday.
The product is available as a research preview to Max subscribers at $200 per month, with priority support for the web version. Each Teammate has an independent desktop, browser, memory, and permission boundaries, and can proactively monitor changes and complete tasks end to end. Further internal data showed that the median employee’s weekly agent hours obtained from Teammates had increased sevenfold from September, and about one-fifth of employees were reaching 40+ agent hours per week.
This setup shifts work organization from one-off prompt sessions to delegated responsibilities and ongoing collaboration. Srinivas explicitly said that the staffing plans and growth logic for the next few years would differ significantly from the traditional model. The pace of agent-scale expansion far exceeds headcount, and the core constraints shift to goal setting, quality judgment, and boundary control.
Original source:
Title: We have about 10,000 teammates (within the last 30 days) in Perplexity's internal workspace right now...
Publisher: Aravind Srinivas (@AravSrinivas)
Date published: 2026-11-05
URL: https://x.com/AravSrinivas/status/2118396712775819430
Source nature: CEO original post
Verification status: The original post and same-day official and media cross-reports were read
Supporting content: about 10,000 Teammates in the internal workspace over the past 30 days, 20 times the number of human employees, about 1,000 active yesterday, and the statement that human-resource planning will change
ABAB AI Insight
Perplexity positions Teammates as persistent colleagues with roles and responsibilities, rather than temporary tools. This represents a clear leap in hierarchy from the early Computer sandbox session model: from a single task and a single context to persistent memory, proactive monitoring, and cross-agent collaboration. The rapid rollout of 10,000 instances within 30 days internally shows that its orchestration layer (multi-model scheduling, permission isolation, Slack closed-loop workflows) already has an engineering foundation that can be replicated at scale, rather than being merely a proof of concept.
The rapid increase in agent hours (sevenfold at the median, with some employees getting an extra full work week) directly changes the marginal cost structure. Traditional software-company expansion depends on hiring, training, and management bandwidth; now, incremental output can be achieved by adding agent instances, while human employees’ main work shifts to setting goals, judging quality, and constraining boundaries. This means that the ceiling for per-capita output is no longer determined by headcount, but by goal clarity and agent reliability.
On the capital path, Perplexity is turning its search-product user base into an entry point for a workflow subscription. The Max plan (about $200/month) carries both Computer and Teammates, forming a continuous set of charging points from information retrieval to task delivery. Running 10,000 agents internally first is both product validation and sales material for proving to enterprise customers that it has deployable digital labor at scale. In the competitive landscape, similar forms such as OpenAI Dots and Anthropic are appearing in parallel, and the differentiators will center on orchestration stability, the granularity of tool permissions, and the reliability of cross-session memory.
Structurally, this is an attempt to shift pricing power from “tool software” toward “digital labor rental.” When the number of agents significantly exceeds the number of human employees, the core variable in organizational design shifts from how many people to hire to how many agents to assign to each role and how to set approval thresholds. If this model is proven transferable, the valuation logic for software companies will partly shift from user growth to the total volume of deliverable agent hours, and human-resource planning itself will become part of the product functionality.
ABAB News · Laws of cognition
Once the number of agents exceeds the number of heads, the expansion bottleneck shifts from hiring to goal clarity and permission design.
The persistent colleague form turns the cost of a single prompt into a subscription for ongoing responsibilities.
Getting large-scale agents working internally first is a necessary prerequisite for selling digital labor externally.