March was the month the big labs started shipping the plumbing for AI agents. Agents have been a demo for a while. This month the tools to build them got a lot more standard.
OpenAI Ships Agent Tools
On March 11, OpenAI released the Responses API and an open-source Agents SDK.¹ The Responses API combines chat completions with built-in tools for web search, file search, and computer use. The Agents SDK handles handoffs between agents, guardrails, and tracing, and it works with models from other providers too.
OpenAI also said the Responses API will eventually replace its Assistants API. If you built on Assistants, start planning the move.
MCP Becomes the Common Connector
This week OpenAI said it will support Anthropic's Model Context Protocol, starting with the Agents SDK, with the ChatGPT desktop app and the Responses API to follow.² MCP is an open standard for connecting models to tools and data. Anthropic released it last November.
This is the part of March I think matters most. When the two biggest model providers back the same protocol, a tool builder can write one integration instead of one per model. That is how ecosystems form. It also gives agents a common way to reach the tools and data they need to do real work.
Designing for Agent Customers
If agents become real users of software, product teams need to design for them:
- API-first: Agents don't click through web pages. Everything a user can do should be reachable through an API.
- Machine-readable docs: Tool descriptions and schemas matter as much as the docs written for people.
- Recoverable errors: Error messages should say what went wrong and what to try next, so an agent can retry without a person.
- Observability: When an agent takes ten steps, you need a trace of each one to debug it.
The last point is the one I think about most. An agent that fails silently is worse than no agent at all.
Still Early
Most agents in production today are narrow: one workflow, a few tools, and a person approving the risky steps. That is the right place to start. Autonomy should grow as teams build the monitoring they need to trust it.
Are you building agents yet, or building products for agents to use? I'd like to hear what's working.