Blog

The raw Messages API tool loop, done properly: schema design, the stop_reason dance, parallel tool calls, and where the loop should terminate.

DSPy optimizes prompts against a metric. Give it a web retriever and the optimizer starts tuning how your agent searches, not just how it answers.

smolagents writes Python to call your tools. That changes how you should design a search tool — and makes multi-step research surprisingly compact.

Use @function_tool to give an Agents SDK agent Serply-backed search, then hand off to a specialist agent that reads the pages.

Write a Serply-backed Haystack component that emits Documents, then drop it into a RAG pipeline alongside your existing retrievers.

Expose Serply's search and scraper endpoints as native Semantic Kernel functions the planner can compose on its own.

Implement BaseRetriever against Serply so your query engines pull from the live web instead of a stale vector index.

Define Serply search as an AI SDK tool, stream the model's reasoning to the browser, and render sources as they arrive.

Pydantic AI validates tool arguments and results at the boundary. Here's how to model Serply's search responses so bad data fails loudly instead of quietly.

Register Serply's search and scraper endpoints once, then let every agent in an AutoGen group chat call them without duplicating tool code.