The new OpenAI Agents Platform

Alessio, the CTO at Decibel, leads a discussion with Spwix, Roman, and Nikunj about the new OpenAI Agents Platform and its API features. The conversation highlights the launch of three built-in tools: a Web Search Tool for enhanced search capabilities, an Improved File Search Tool for data uploads, and a Computer Use Tool for developers. The new Responses API replaces the previous chat completions API, supporting longer, multi-turn tasks and agentic workflows. Concerns regarding the Responses API are addressed, noting that the chat completion API will remain available as a foundational tool. The migration process to the Responses API is simplified, ensuring a smooth transition for users of the assistance API, which is set to sunset in 2026.

The introduction of the Web Search feature within the Responses API showcases its accuracy improvements over previous models, focusing on accurate information retrieval and citation. The integration of web search with other platform features allows for real-time data structuring. Questions arise regarding publisher control over their appearance in the web search tool and the implications of knowledge cutoffs in relation to live data access. The importance of managing search costs through parameters like similarity cutoffs is discussed, along with the potential for creative solutions in search query budgeting.

A query is processed using a file search tool to retrieve a user's preferences, followed by a web search for related information, resulting in a curated list. This integration of file search and web search is executed through a single API call, allowing developers to creatively combine these tools for enhanced user experiences. The file search capability has expanded to include more file types, query optimization, and custom re-ranking, with metadata filtering being a highly requested feature. Companies may prefer not to rebuild complex competencies like embeddings and chunking in-house.

The computer use model allows developers to build agents that can complete tasks using a computer or browser, with potential for multi-turn actions. The merging of various models into the main branch is discussed, with comparisons made to the introduction of vision capabilities in GPT-4. The Agents SDK is introduced, highlighting new features such as support for types and guard railing, facilitating parallel execution and action blocking. Tracing support is added, allowing users to view traces in the OpenAI dashboard and integrate with other tracing providers.

Nikuj points out the advantage for developers in separating logic rather than relying on a single agent for multiple tool calls. A triage agent is introduced to direct intents to different agents, and new user interface logs on the OpenAI dashboard are announced for better troubleshooting of workflows. The agents SDK and tracing UIs are expected to enhance agent workflows for users and developers, with a commitment to ease workflows and excitement about future developments.