Perplexity's Sonar model update adds structured tool-calling and multi-step reasoning to what was previously a search-and-summarize engine. Users can now ask Sonar to perform sequences of research steps, compare sources, and produce structured outputs. This is a meaningful step from retrieval toward agency.
AGNT occupies the complementary position in the stack. Search engines — whether Google, Perplexity, or any RAG pipeline — are excellent at finding information. They are not designed to execute commerce transactions: booking a table, confirming availability, processing payments, sending reminders. That is AGNT's domain.
The composition is natural. A Perplexity-powered agent researches restaurants in Seminyak with ocean views. It finds options, reads reviews, compares menus. When the user says 'book the second one for Friday at 7pm,' the agent needs a commerce layer. AGNT's MCP tools handle the booking: check availability via the venue's Sam agent, confirm the reservation, send a WhatsApp confirmation, and schedule a reminder.
We tested this flow using Perplexity's API for the research phase and AGNT's /mcp/sse endpoint for the execution phase. The handoff is clean because both sides speak structured tool schemas. Perplexity returns structured venue data. AGNT accepts structured booking requests. No custom middleware required.
The broader pattern: search-first AI and action-first AI are converging. Perplexity is adding action capabilities. AGNT is adding richer context retrieval. The meeting point is protocol-level interop — MCP, A2A, and structured tool schemas that let each system do what it does best without trying to replace the other.