KM4BI

Does Your Computer Belong to Codex? I Went to OpenAI to Ask

At OpenAI, agents are moving beyond chat: they navigate software, assemble context, and reshape knowledge work around direction, judgment, and iteration.

Published

A personal computer is starting to serve two operators at once. The human states an objective, adds context, and steps in when needed; the agent moves through windows, tests interfaces, reads documents, and completes the work. In Nate B. Jones’s conversation with Andrew and Ash at OpenAI, that uneasy question — whose computer is it now? — becomes a practical account of how knowledge work is changing.

Computer use fills the gaps

Dedicated connectors, APIs, and MCP can be faster and more token-efficient. They are still worth building. Yet many real workflows pass through government forms, PDFs, old websites, and services that will never expose a clean machine interface. Computer use acts as the universal fallback.

The capability was once an impressive but slow demonstration. The guests say it has become fast enough that an agent can move from a connector to visual interaction for an uncovered step without drawing attention to the switch. Product design changes with that maturity: mainstream users should not need to inspect scripts and selectors, but they still need enough visibility to trust what the agent is doing.

Adoption follows context, not job labels

Software development tipped early because its context was already available through local files and command-line tools. Legal work followed where large document collections were accessible. As connectors and computer use improved, more functions crossed the same threshold.

That pattern weakens the old split between technical and nontechnical roles. Finance and business teams at OpenAI are described as replacing some spreadsheet workflows with shareable, evolving sites that behave like internal applications. They are effectively becoming developers, even though tooling such as version control has not yet caught up with their new role.

Abundant output raises the value of judgment

The guests expect work products — sites, tools, documents, and other AI-native artifacts — to keep improving in quality, intent recognition, context retrieval, and sourcing. As those capabilities rise, the constraint moves upstream: deciding what to make, testing whether it is useful, and refining it with other people.

Easy creation also creates a convergence problem. An organization can produce many plausible things at once, but still needs a focused and coherent product. Taste, clarity, and the willingness to discard output become central skills. Removing busywork does not remove human responsibility; it exposes it.

Collaboration continues during the run

Long-running agents do not fit a turn-by-turn chat model. A user may see the system heading in the wrong direction, or realize twenty minutes later that the original instruction was mistaken. Interruption and redirection therefore need to feel natural throughout execution.

Clarifying questions are part of the same design. One timely yes-or-no question delivered to another device could unblock substantial work without asking someone to sit at a desk. The goal is not constant attention, but efficient moments of human judgment.

Voice in, visuals out

Voice is described as the fastest way to provide rich context, while reading is faster than listening. That points toward a mixed interface: users speak at conversational speed, then scan generated visuals and answer focused questions. Early adopters already work this way, but the behavior has not yet reached most users.

Proactivity is a harder extension. An agent might surface a retention blind spot before anyone asks, but unsolicited analysis must clear a higher bar for accuracy, relevance, and cost. Running inference for hours to deliver weak observations would turn assistance into expensive spam.

From personal wins to team systems

The recommended habit is straightforward: before doing a task manually, try giving it to the agent. Team-level gains appear when an early adopter turns a repeated problem into something others can use — an internal site, a reporting workflow, or an automation that publishes actionable issues. Concrete shared tools spread fluency faster than abstract training because colleagues receive value while seeing what is possible.

Source