OpenAI president Greg Brockman says developers are "retooling the world" to make software accessible to AI agents through purpose-built connectors, but computer use—where agents operate interfaces the same way humans do—could offer a simpler alternative. In a conversation on the a16z podcast this week with hosts Ben Horowitz and Erik Torenberg, Brockman argued that the industry's current approach of building MCP servers, command-line interfaces, APIs, and other integrations for every application places an unnecessary burden on developers. According to Brockman, OpenAI has been considering computer use since its earliest days, and he believes it could help the field advance without requiring custom integrations at every turn.
The concept dates back to November 2015, when OpenAI's team discussed using reinforcement learning directly against a computer interface—screen pixels, keyboard, and mouse—the same interface humans use. Brockman told the podcast hosts that the company laid out a three-step plan during an offsite meeting that month, which the team has largely followed for the past decade. He views this approach as potentially unlocking the computer for almost any task, allowing agents to interact with software without developers needing to construct and maintain numerous specific connectors. Brockman considers this capability part of why GPT-6 Astra, OpenAI's newest flagship model, is "pretty reasonable to call" AGI.
Brockman described the current wave of connector infrastructure as creating access "in this almost stilted way that is not really meant for humans," according to the podcast. He emphasized that agentic use cases fundamentally depend on tools, which explains why MCP servers and similar connectors have proliferated. But he questioned whether a different path might exist: "What if it's more behaving like a human? Can it just use a computer?" Brockman pointed to everyday tasks—clicking through menus, entering data into spreadsheets—as activities people shouldn't be doing themselves, suggesting agents could handle them through direct computer interaction rather than specialized integrations.
Yet the industry hasn't abandoned structured integrations. AWS recently launched a managed consent portal and session binding endpoint for AgentCore Gateway, a feature of Amazon Bedrock AgentCore that links agents to external tools and services, with a detailed walkthrough published Monday. The portal lets administrators configure targets like GitHub and Slack, then send developers a URL where they can sign in and connect services independently. Even at OpenAI, plugins remain the preferred method for Codex to work directly with services such as Slack, Gmail, and GitHub, despite the Chrome extension released in May that allows agents to operate within live browser sessions. Brockman positioned simplicity as the guiding principle—"the North Star"—for AI development, and reliable computer use could represent one route to reach it. For now, though, connectors continue to serve as the industry's primary mechanism for expanding agent capabilities. The tension between building for machine-readable integrations versus teaching machines to use human-readable interfaces will likely define how much infrastructure overhead the next generation of AI deployment demands. Organizations betting heavily on either path face meaningful architectural commitments that won't reverse quickly or cheaply.