OpenAI's Agents SDK and ACP: How OpenAI Is Competing in the Agent Infrastructure Race

OpenAI's 2026 Agent Offensive

While Google dominated headlines with I/O 2026, OpenAI shipped three major releases: the Agents SDK update with native sandbox execution, the Codex MCP server for multi-agent pipelines, and the Agentic Commerce Protocol (ACP) for ChatGPT shopping.

Agents SDK and Codex Updates

The Agents SDK introduces SandboxAgent with isolated workspaces, file inspection, and command execution. Multi-agent orchestration with hand-offs, guardrails, and traces is built in. Codex CLI now works as an MCP server, enabling multi-agent pipelines where a designer agent hands off to a developer agent, which hands off to a tester agent.

Codex has evolved from autocomplete into a four-surface system (Desktop, CLI, IDE, Cloud) with GPT-5.5, 1M token context, and pay-as-you-go pricing.

ACP and the Three-Way Platform Race

OpenAI's Agentic Commerce Protocol competes with Google's UCP. ACP ingests structured catalog data and surfaces products in ChatGPT. The three-way race between Google (Antigravity + A2A/UCP/AP2), Anthropic (Managed Agents + MCP), and OpenAI (Agents SDK + ACP) is a net positive — all three open their protocols and SDKs, creating an interoperable ecosystem that open-source projects can build on.