Perplexity Is Changing How AI Apps Search the Web and Handle Citations

Perplexity is making an important change to the way developers build AI applications using its platform. The company set September 27, 2026 as the retirement date for its Sonar Chat Completions endpoints, directing developers toward the newer Agent API instead. But this is not simply a case of replacing one API address with another. The new architecture changes how web search is activated, where citation information appears and how applications need to handle the different stages of an AI agent’s work. For developers building research tools, AI search products or assistants that depend on current information from the web, those differences matter.

Sonar is giving way to the Agent API

Perplexity’s older Sonar integration used the familiar chat-completions model. Applications could send a conversation to the /chat/completions endpoint and receive a generated response with web-grounded information and citation data. The Agent API uses a broader architecture. Instead of treating the request primarily as a conversation with a search-powered model, Perplexity now treats it as an agent task that can involve different capabilities during execution. Those capabilities can include web search, multi-step research, code execution and other tools. The result is a platform designed for more complex AI workflows, but it also means developers need to be more explicit about what they expect the agent to do.

Web search is no longer something developers should assume

One of the biggest differences is the way web search works. A basic Agent API request does not automatically mean that Perplexity will search the web. Developers can provide the web_search tool or use a preset that includes search capabilities, but simply making a search tool available does not always mean the model will decide to use it. That introduces an important distinction. An application can send a question to an AI model and receive an answer based on the model’s existing knowledge, or it can instruct the agent to search the web and build its answer using current external information. For applications where fresh information is essential, developers may need to explicitly require web search instead of leaving the choice entirely to the model.

Why explicit search matters

This change may seem technical, but it can significantly affect the information an AI application returns. Imagine two applications receiving exactly the same question. One application allows the agent to answer without searching, while the other requires web search before generating its response. Although both may use Perplexity’s Agent API, the resulting answers could be based on very different information. The first might rely primarily on the model’s existing knowledge, while the second can retrieve recent sources from the web. That makes search configuration an important part of how developers design AI products, and it also means that the name of the AI model alone no longer tells you whether a response was actually grounded in live web information.

Citations have moved inside the agent output

Another major change affects citations. Under the older Sonar response format, developers could access citation and search-result information through familiar top-level response fields. The Agent API organizes the response differently. Search results are now represented as output generated during the agent’s execution, meaning developers need to inspect the response’s output data and find the search-result item created when the web-search tool runs. This reflects the larger philosophy behind an agent system. A response is no longer necessarily one block of generated text accompanied by a few pieces of metadata. Generated text is one output, web search can create another and additional tools can generate still more.

Apps that keep the old parser could appear to lose citations

The response change creates an important migration risk. A developer could successfully move an application to the Agent API while leaving the old citation-parsing code largely unchanged. The AI request might work and web search might also run successfully, but the application could still display no sources because it is looking for citation information in the wrong part of the response. Perplexity’s migration guidance warns developers not to rely on the old top-level citation fields when working with Agent API responses. For products that display references to users, store research evidence or measure how frequently websites appear in AI answers, updating the response parser is therefore an essential part of the migration.

The Agent API separates reasoning from retrieval

The architecture also reveals something larger about the direction of AI systems. With Sonar, web search and the model felt closely connected. With the Agent API, the model can increasingly be viewed as the reasoning component while search becomes one capability the agent can choose or be instructed to use. That separation gives developers more flexibility. An AI agent might search the web for one part of a task, analyze the retrieved information, run code for another part and then produce a final response. Search is therefore no longer necessarily the entire experience. It can become one step inside a much larger workflow.

Developers get more control, but also more responsibility

This flexibility comes with additional decisions. Developers now have to think about when current information is required, when model knowledge is sufficient and when the application should force a search. Leaving the decision to the agent may reduce unnecessary search operations in some applications, but products dealing with breaking news, current product information, recent research or frequently changing data may need much stricter rules. If web search is essential to the product, relying entirely on the model to decide whether to search may produce inconsistent behavior.

Old Sonar parameters can also cause migration problems

Moving to the new API requires more than changing the endpoint. Perplexity warns that the Agent API validates request fields more strictly than the previous Sonar interface. Parameters left over from older Sonar integrations can cause requests to return an error if the Agent API does not recognize them. That means developers should review the complete request structure rather than copying an existing Sonar request and replacing only the URL. Search settings, model configuration, token controls, tool definitions and response handling may all need to be checked against the new API format.

Successful HTTP requests do not always mean a successful agent run

Agent systems also introduce more complex execution states. Traditional API integrations often treat a successful HTTP response as confirmation that the operation itself completed correctly. With Perplexity’s Agent API, developers also need to examine the status of the agent response. A run can have states such as completed, failed, incomplete or cancelled. That means robust applications need to understand not only whether the server responded, but whether the actual agent workflow finished successfully. This becomes even more important for longer tasks involving several tools or multiple stages of research.

Streaming is becoming more complex too

Applications that stream AI responses in real time also need to account for the new architecture. A simple chatbot can often display text as tokens arrive and stop once the response finishes. An agent may have several different terminal states because the task itself can involve more complicated execution. Developers therefore need to handle successful completion as well as incomplete, failed or cancelled runs. This is another sign that AI APIs are evolving from simple text-generation interfaces into platforms capable of managing longer-running software processes.

The change could affect how websites appear inside AI products

The new architecture is also interesting for publishers. A website cannot become part of an AI application’s retrieved evidence if the application never searches the web in the first place. Under the Agent API, that decision can happen before a search system ever evaluates which websites are relevant. If developers make web search optional, two users asking similar questions could potentially see different levels of web grounding depending on whether the agent decides external information is necessary. That makes AI visibility increasingly dependent on the configuration of the application itself, not only on whether a page is accessible or useful.

Missing citations may not mean a website disappeared

This distinction could become particularly important for companies monitoring how often their websites appear in AI-generated answers. A sudden reduction in citations does not necessarily mean the content has become less relevant. The application may have stopped requiring web search, the agent may have decided not to call the search tool or search may have run while the application failed to read the new result structure correctly. Those are very different situations, so understanding how an AI application is configured is becoming increasingly important when interpreting changes in citation visibility.

Perplexity is turning its API into a broader agent platform

The move away from Sonar Chat Completions reflects a wider transformation taking place across the AI industry. Early generative AI APIs were largely built around a simple pattern: send text to a model and receive text back. Agent platforms are much more ambitious. They allow models to interact with tools, search external information, execute code and work through tasks containing multiple steps. Perplexity’s Agent API follows that direction by combining model access with web search and additional capabilities inside one orchestration system.

Search becomes one tool among many

Perhaps the most important change is conceptual. Web search is no longer something developers should automatically associate with every Perplexity AI request. It is becoming one tool available to an agent. That means model selection and search behavior can increasingly become separate decisions. A developer can choose which model should reason about a task while independently deciding whether the agent should access current web information. For applications where citations and fresh information are essential, that distinction needs to be reflected directly in the product’s architecture.

A bigger change than an endpoint migration

Perplexity’s move from Sonar Chat Completions to the Agent API may initially look like a routine API upgrade. In practice, it represents a broader change in how AI applications are structured. Developers must now think more carefully about when web search happens, how retrieved information moves through an agent workflow and where evidence appears in the final response.

The Agent API provides significantly more flexibility than a traditional chat-completions interface, but that flexibility also makes application design more important. As AI products continue moving toward agents capable of using multiple tools and completing longer tasks, this type of architecture is likely to become increasingly common. The model may generate the answer, but the application increasingly determines which tools, information and external sources the model is able to use along the way.