Agent API
A model invocation interface designed for agent scenarios, offering tool calling, multi-step planning and structured output. It enables developers to integrate external tools and build autonomous task pipelines.
Timeline (last 90 days)
Version 5.0.0 fully migrates language generation capabilities from Perplexity's Sonar Chat Completions API to the Agent API.
Version 5.0.0 replaces the original Sonar model IDs and provider options with the Agent API's presets, models, and tools.
The new API modifies the data structures for request and response metadata, raw stream events, usage, and cost data.
The shift from the Chat Completions paradigm to the Agent API reflects Perplexity's transition from conversational completion toward Agent-centric callable tool orchestration capabilities.
The new version is able to recover missing text (including incomplete responses) from the Agent API's terminal events and output items, while avoiding duplication of content already received incrementally.
The SDK can now accept the Agent API's native tool traces (such as financial data results) without erroneously validating them as web search results.
The new version retains URL citation annotations as sources in the Agent API's streaming output, ensures each source URL is emitted only once, and preserves search result IDs for citation association.
Agent API clients can no longer directly control the detailed behavior of the underlying model, as the trade-off for gaining the platform's ability to flexibly combine retrieval tools.
All Facts (8)
Version 5.0.0 fully migrates language generation capabilities from Perplexity's Sonar Chat Completions API to the Agent API.
50%UnverifiedVersion 5.0.0 replaces the original Sonar model IDs and provider options with the Agent API's presets, models, and tools.
50%UnverifiedThe new API modifies the data structures for request and response metadata, raw stream events, usage, and cost data.
50%UnverifiedThe shift from the Chat Completions paradigm to the Agent API reflects Perplexity's transition from conversational completion toward Agent-centric callable tool orchestration capabilities.
50%UnverifiedThe new version is able to recover missing text (including incomplete responses) from the Agent API's terminal events and output items, while avoiding duplication of content already received incrementally.
50%UnverifiedThe SDK can now accept the Agent API's native tool traces (such as financial data results) without erroneously validating them as web search results.
50%UnverifiedThe new version retains URL citation annotations as sources in the Agent API's streaming output, ensures each source URL is emitted only once, and preserves search result IDs for citation association.
50%UnverifiedAgent API clients can no longer directly control the detailed behavior of the underlying model, as the trade-off for gaining the platform's ability to flexibly combine retrieval tools.
50%