Tools

Perplexity Teammates: AI Agents Take On Long-Running Work

Perplexity Teammates extends AI search into multi-agent collaboration and long-running tasks. This analysis examines its context management, tool use, error recovery, and completion criteria, and compares it with frameworks such as LangGraph.

Font size

Perplexity has released Teammates, which lets multiple AI Agents collaborate on complex, continuous work. Its implications are direct: what users assign to AI may no longer be a single question, but a task that requires tracking and validation.

Perplexity Teammates: A Deep Dive into AI Agents for Long-Running Tasks

Perplexity describes Teammates in straightforward terms:

“Teammates are agents for complex, ongoing work.”

It does more than answer a question once. After a user submits a goal, multiple Agents divide the work, make progress in parallel, and maintain the task state.

This differs from earlier Perplexity use cases. The typical pattern was to ask a question, run a search, and receive an answer with citations. Teammates instead assigns a goal to multiple independent Agents that collaborate to complete a chain of work.

The important development is not another chat box. It is that the product boundary for AI Agents is beginning to shift toward long-running tasks.

Teammates Is Not Another Chat Box

The endpoint of a one-off question is an answer. The endpoint of a long-running task is a deliverable. The distance between them may span several hours and a dozen steps.

For example, “research the US energy storage market and prepare a briefing for an investment committee” does not fit neatly into a single Prompt. The process usually involves:

  • Collecting company filings, industry reports, and regulatory documents;
  • Cross-checking data that uses different definitions or methodologies;
  • Saving sources already found to avoid repeated searches;
  • Adjusting the task when evidence is missing rather than starting over;
  • Compiling the findings into a cited document.

Teammates is designed around multiple independent Agents collaborating toward that goal. Each Agent handles a different subtask, shares the overall objective, and maintains its own working context.

The core unit in the interface is therefore no longer merely a question and answer. It is closer to a task.

Users also have to answer a different set of questions:

  • What is the goal?
  • How far should the task progress?
  • Which sources are allowed?
  • Which actions require confirmation?
  • What counts as complete?

This is the most practical difference between Teammates and an ordinary chat box. A chat box revolves around messages; Teammates revolves around tasks.

From Finding Answers to Assigning Work

Perplexity has traditionally centered its product on AI search. Users enter a query, the system retrieves and organizes web content, and the response includes citations.

Teammates extends that pipeline one step further. Search is no longer only the final output; it becomes one tool Agents use while completing the work.

The interaction model changes as well.

In conventional search, the user breaks the problem into questions. With Teammates, the user acts more like a project lead: they define the goal and let multiple Agents divide, coordinate, and track the work.

Open-ended search usually requires a person to decide what to look up next. A sustained task requires the system to maintain its own plan.

Perplexity is moving from an AI search engine toward an Agent platform that can execute work over time. The key point is not that a search company has released an Agent. Perplexity already had two capabilities that support this direction:

  1. Finding information across public web pages;
  2. Attaching sources to conclusions.

Most general-purpose Agent platforms offer search tools, but they do not necessarily make retrieval, citation, and verification part of the default workflow. That is where Perplexity's experience could provide an advantage.

Long-Running Tasks Are the Real Test

The first ten seconds of an Agent demo are usually the easiest part to impress.

It decomposes the goal, invokes tools, and produces a document that appears complete. The real pressure comes later: Does the state survive after two hours of execution? Can it recover from a failed tool call? Will it revisit sources when they become outdated?

Long-running tasks require at least five capabilities.

1. Context Management

The system knows the goal when the task begins. After several hours, it must still remember the constraints, completed work, and current blockers.

This is not simply a matter of enlarging the context window. A larger window holds more text; it does not guarantee that the Agent will retrieve the relevant material correctly.

2. Tool Use

Search is only one tool. Sustained work may also involve reading documents, running code, calling APIs, and writing intermediate results back into the task workspace.

If a tool returns no results, times out, or changes its response structure, the Agent cannot simply continue with text that appears plausible. It should recognize the failure and then choose whether to retry, take another path, or return the issue to the user.

3. Task Handoffs

Handoffs are where multi-Agent collaboration most often breaks down.

One Agent may produce a market analysis while another verifies the financials. The data definitions, units, and time ranges used by the first Agent must carry over in full. Otherwise, multi-Agent execution merely amplifies the same error in parallel.

4. Error Recovery

In real-world continuous work, success rarely comes on the first attempt. Goals change, web pages disappear, sources conflict, and APIs impose rate limits.

An Agent that can reliably finish a workflow needs to know which errors are retryable, which require a change of plan, and which must stop execution and bring in a human.

5. Completion Criteria

The most dangerous Agent failure is not saying “I cannot do this.” It is declaring the task complete after completing only the surface actions.

For example, a user asks for a cited industry report. The Agent generates the document but never opens each citation to verify it. That omission may pass unnoticed in a short answer, but in a long-running task it becomes a delivery failure.

The meaningful evaluation metrics for Teammates should therefore be task completion rate, recovery rate after retries, number of human interventions, and citation accuracy—not a three-minute demo.

The Product Questions Behind 47 Points on HN

The announcement reached the Hacker News front page, receiving 47 points and 36 comments.

The response was not a breakout hit, but it was enough to bring the product questions into view among developers. HN page:

https://news.ycombinator.com/item?id=50395866

Developers tend to ask several questions.

First, how does Teammates' execution model differ from frameworks such as LangGraph, CrewAI, and AutoGen?

Second, can users export task history, context, and intermediate artifacts?

Third, can users control which data sources, tools, and permissions the Agents use?

Fourth, can an interrupted task resume, or must the user start it again?

Teammates currently looks more like an end-user productivity product than a framework for developers to orchestrate state graphs. Existing Agent frameworks focus on nodes, state, conditional branches, and human approvals. Teammates focuses on assigning goals, coordinating multiple Agents, and tracking progress.

Both approaches may be viable.

Developers need control. Business users need less configuration and clearer results. The platform that can provide both has a stronger chance of becoming the default entry point for long-running work.

A Search Company's Path to Agents

Perplexity's decision to build Teammates follows a logical product trajectory.

Search engines solve the problem of finding information. Teammates aims to solve the problem of completing work around that information. The addition is not a chat window but an execution layer: task planning, Agent division of labor, tool use, state persistence, source attribution, and final delivery.

Compared with general-purpose Agent platforms, it may have an advantage in three areas.

First, search and citations are native capabilities.

Users do not need to install a search plug-in and then wait for an Agent to assemble its own research process. For tasks centered on public information, such as market analysis, competitor research, and policy searches, this is a direct advantage.

Second, the product entry point is lighter.

General-purpose frameworks often require developers to design the Agents, state, and workflow before any work can begin. A Teammates user only needs to describe the goal. The tradeoff is a coarser degree of control.

Third, results are easier to explain.

Perplexity has long emphasized citations. If every conclusion from a long-running Agent can be traced to a source, users will be more willing to delegate work to it.

Its limitations are equally clear.

General-purpose Agent platforms can connect to internal enterprise data, private APIs, and custom tools. If Teammates relies mainly on public search, it will struggle with work that requires access to internal systems. Code changes, operational configuration, and cross-SaaS approvals all demand deeper permissions and execution capabilities.

There is another problem: search is good at collecting currently accessible information, but it is not inherently good at determining whether a workflow is correct.

That is the layer an Agent platform ultimately has to add.

Editorial Perspective

Teammates matters because it signals a shift in how people enter the Agent workflow: from a temporary assistant to a continuous collaborator. Users used to organize work around a single conversation. Task state, failure recovery, and delivery validation now converge in one interface. That changes how users evaluate AI. They no longer judge only whether a response sounds fluent; they assess whether the task finished, whether the citations can be verified, and whether execution can continue after a failure.

My assessment is that search-based platforms will enter research, due diligence, and information organization first, rather than immediately capturing the entire enterprise automation market. The former depends on public sources, citations, and rapid delegation, all of which align with Perplexity's existing strengths. The latter requires internal permissions, long-term memory, reliable recovery, and human approvals.

Unlike earlier Agent releases, Teammates packages multi-Agent collaboration directly as a user-facing product rather than offering only framework capabilities. Its 47 points and 36 comments show that it has entered developers' awareness, but they do not mean the product is settled. Its actual completion rate, permission model, and task portability remain unknown. Those metrics will determine whether Teammates is a successful demonstration or infrastructure for everyday work.

Sources

Source: https://www.perplexity.ai/en-GB/hub/blog/teammates-are-agents-for-complex-ongoing-work