Perplexity is giving developers a new way to shape its AI agents around the work they actually need done.

The company has added Skills to the Perplexity Agent API, allowing developers to combine Perplexity’s built-in capabilities with their own reusable instructions and workflows. The update was announced by Perplexity CEO Aravind Srinivas and the company’s developer account on July 17.

“Agents aren’t defined by a single system prompt,” Perplexity Developers wrote in the announcement. “They are assembled from many capabilities that developers extend and compose.”

That distinction matters. For much of the recent AI boom, building an agent has meant writing an increasingly large system prompt, connecting a few tools and hoping the model consistently follows the intended process. Custom skills offer a more modular approach: instead of placing every instruction in one long block of text, developers can package specialized knowledge and procedures into separate capabilities that an agent can use when needed.

Perplexity’s example is straightforward. A developer could pair the company’s built-in office or PDF skill with an inline design skill created specifically for a particular workflow. One capability might handle extracting information from a document, while another determines how that information should be structured, branded or presented.

That sounds like a small change, but it reflects a broader shift in how AI applications are being built. The most useful agents are rarely general-purpose systems operating from a single prompt. They are collections of focused abilities: searching the web, reading files, querying business data, formatting reports, checking calculations or following an organization’s internal process.

Perplexity’s Agent API already provides a foundation for that kind of system. The API gives developers access to multiple model providers through a unified interface, along with tools such as web search, URL fetching, finance search and people search. Developers can also use presets designed for different kinds of work, including faster responses, more advanced search and deeper research workflows. Perplexity describes the Agent API as a unified interface for models, tools and reasoning controls.

Custom skills extend that foundation beyond the tools Perplexity supplies out of the box. A company could, for example, build a skill that explains how its support team handles refunds, how its analysts prepare weekly market notes or how its engineers review a particular type of code. The agent would then have access to a repeatable workflow instead of relying on a fresh interpretation of the same instructions every time.

That could make agents easier to maintain, too. When every behavior lives inside one system prompt, changing one part of the workflow can create unexpected effects elsewhere. Modular skills give developers a clearer way to update a specific capability without rewriting the entire agent.

The idea is already becoming familiar across the AI industry. Perplexity’s own Computer product has promoted skills as reusable instruction sets for tasks such as research, presentations, data analysis and chart creation. Its documentation says skills can work together, with one capability gathering information and another turning the results into a report or presentation. Users can also create their own skills for recurring workflows. Perplexity’s guidance on Computer skills describes them as specialized playbooks that can be activated and combined.

Bringing that concept to the Agent API is significant because it moves skills from a consumer-facing assistant feature into the developer platform. Instead of only using Perplexity’s prebuilt workflows, companies can begin designing their own agent capabilities and embedding them into applications.

There is also a practical advantage for teams working across different models. Perplexity’s Agent API is designed to support models from providers including OpenAI, Anthropic, Google, xAI (now SpaceXAI) and Perplexity itself. A skill can provide the workflow and operating instructions while developers retain flexibility over which model handles the task. That separation could make it easier to change models without rebuilding an entire application from scratch.

Still, custom skills do not eliminate the hard parts of agent development. Developers will need to decide when a skill should activate, what information it can access and how its output should be checked. A poorly written skill can create the same problems as a poorly written system prompt—only now the behavior may be distributed across several components.

Observability will be important as well. When an agent combines multiple skills, tools and models, a failed result may be difficult to diagnose. Developers will want to know which skill ran, what instructions it received, which tools it called and where the workflow went off course. The more capable these systems become, the more important it is to make their decisions traceable.

Perplexity’s announcement does not yet spell out every implementation detail of the new feature, including the full format for creating and managing skills through the API. But the direction is clear. The company is positioning agents as assembled systems rather than monolithic prompts.

That is a more realistic picture of how useful AI software will likely be built. A travel agent may need a flight-search skill, a policy-checking skill and a budgeting skill. An internal research assistant may combine web search with company-specific reporting rules. A design tool may need separate capabilities for interpreting a brief, generating assets and checking whether the final result follows brand guidelines.

The promise of custom skills is not simply that agents can do more. It is that developers can give them a clearer division of labor.

For Perplexity, the update also broadens the appeal of the Agent API beyond applications that merely add web search to a chatbot. The company is offering a managed environment where developers can combine models, real-time information, built-in tools and their own specialized workflows.

The next test will be whether those skills are dependable enough for production use. If they are, the update could help turn the Agent API from a flexible model-and-tools interface into something closer to an operating layer for customized AI workers.