OpenAI's Agents SDK and Responses API

Agents just got a whole lot easier to build

Spring has sprung, and building agents is getting easier by the day. Last week, we talked about Anthropic’s Model Context Protocol, or as Anthropic calls it, the USB-C port for AI. OpenAI just launched their new Agents SDK, alongside the new Responses endpoint this week.

Having these both as tools in your agentic toolbox drastically reduces the time it takes to build multi-agent systems.

OpenAI’s Agents SDK and Responses API

First, just to clarify, the SDK and new endpoint are separate entities, but the Agents SDK defaults to using the Responses API. Here are the highlights for each:

Responses API

• Very similar to Chat Completions API endpoint

• Three built-in tools: Web search (powers ChatGPT search), file search (retrieve files from your vector store in OpenAI), Computer Use (Agent interacts with your computer = OpenAI operator)

• messages appears to have been replaced with input

• System messages/developer messages can be passed in the input array but you can also pass instructions: “Inserts a system (or developer) message as the first item in the model's context.” Taking the system instructions outside of the messages array/input is essentially how Anthropic does it as well.

Using both instructions and a system message

Agents SDK

• Helps orchestrate multiple agents

• Handoffs: Agents can delegate specific tasks to specialized agents. For example, a customer support agent can escalate a billing issue to a billing agent. So rather than having a single agent manage everything from billing to general support, you can have specialized agents and a triage agent that directs queries to the right agent.

• Guardrails: Developers can enforce constraints on agent behavior. For example, a customer support agent can be restricted from making any sort of political statement.

• Observability & Debugging: The SDK includes built-in tracing tools to help monitor traces

Time to build!

Between the new Agents SDK which helps with orchestration and Anthropic’s MCP which helps with retrieving the right context, building agents has never been easier. We plan on doing a deeper dive into MCP in the coming weeks as we look at integrating a few servers into PromptHub. In the meantime, final plug for PromptLab, our free 3-day webinar starting tomorrow - hope to see you there!