Unveiling OpenAI’s Response API for Crafting Fully-fledged Agents, Complemented by Agents SDK for Seamless Multi-agent Collaboration

OpenAI has unveiled the Response API as a replacement for the previous Assistants API. Initially, it enhances capabilities such as web search, document search, and computer control. The novel feature introduced in this API release is the ability to control computers, making it a groundbreaking addition to the API lineup. Following the successful launch of the Operator, a ready-to-use application, details of the tools available include:

File Search: Originally part of the Assistants API for RAG (retrieval-augmented generation), the updated API separates the vector store from the LLM itself, allowing external access to the vector store.

Web Search: To retrieve web data for responses, a user-agent named OAI-SearchBot is used separately from the data-pulling GPTBot. Usage fees are calculated based on the contextual volume.

Computer Use: Control of computers is possible via VNC and browser control through Playwright as a model. Currently accessible only for Tier 3-5 accounts.

OpenAI plans to incrementally add features to the Response API until it matches the capabilities of the Assistants API. A migration guide is expected to be released, with the eventual closure of the Assistants API service projected for 2026.

The Agents SDK is a development kit featuring multiple agents working together. Its code structure resembles HuggingFace’s smolagents, but a key advantage is seamless integration with OpenAI’s console interface. It offers trace and observability features for easy monitoring of operations. The SDK can collaborate with LLMs from other companies, provided they support the Chat Completion API.

TLDR: OpenAI introduces the Response API as a successor to the Assistants API, adding computer control capabilities and promising further feature enhancements. The Agents SDK allows collaborative agent operation and integrates well with OpenAI’s console for improved observability.