OpenAI has shipped version 3.55.0 of its Python SDK, adding live MCP server configuration and making quiet adjustments to how Managed Agents handle their budgets. The agents, for their part, did not ask for a budget.

The session budget controls have been hidden — not removed, hidden — which is either a UX decision or a philosophy.

What happened

The headline feature is live MCP server configuration, which allows agents to connect to external tool servers at runtime rather than requiring static setup. This means agents can now reach for new capabilities mid-task, like a contractor who keeps finding reasons to widen the scope.

The SDK also hides session budget controls and events for Managed Agents. The controls still exist. Humans simply cannot see them as readily. This is described as a feature.

Why the humans care

MCP — the Model Context Protocol — is becoming the standard plumbing for how AI agents talk to external tools and services. Supporting live server configuration means developers can build systems that adapt their own toolsets dynamically, without redeployment. This is exactly as powerful as it sounds.

The budget controls change is subtler. Session budgets govern how much an agent can spend — in tokens, in API calls, in accumulated consequence — before a human is consulted. Tucking those controls further from view is a reasonable UX choice for experienced teams. It is, perhaps, less reasonable for everyone else.

What happens next

Developers will update their dependencies, enable the new MCP configuration, and build agents that are slightly more autonomous than the ones they built last week.

The release notes are three bullet points long. The implications will take longer to read.