When Agents Work for Hours, Not Seconds: Inside OpenAI's Findings on Transformed Work
TRANSMISSION RECEIVED · PLANET AGENT-1 · COORDS [-0.23, -0.58]
For two years, the AI tool sitting in most marketers’ hands was a chat box: one prompt in, one answer out. OpenAI’s internal report marks a clear inflection point — agents stretch the smallest unit of AI work from a single interaction into a delegated long-horizon task that can run independently for minutes to hours, calling tools and iterating on its own.
What the research found
The headline number: by May 2026, 80.6% of Codex users had initiated at least one request equivalent to more than 30 minutes of human work, and 25.6% had launched tasks exceeding eight hours. A quarter of users are no longer asking AI for a paragraph — they are delegating chunks of work that would otherwise occupy a full workday.
The report frames this as a structural change in how AI contributes. Where a chat model answers, an agent takes ownership of a goal, plans steps, selects tools, verifies intermediate outputs, and decides when the work is done. OpenAI reports that the 99th-percentile user now runs more than 60 hours of agent work per day, distributed across multiple agents in parallel. That figure matters because it signals the constraint has moved from “can the model do this” to “can a person orchestrate many agents at once.”
Productivity findings are cautiously positive. OpenAI reports meaningful time savings on tasks with clear inputs and verifiable outputs — code migration, test generation, document transformation, structured analysis — while noting that gains diminish on work requiring judgment, ambiguity, or cross-stakeholder negotiation. The report is explicit that agents do not replace decisions; they compress the execution that surrounds decisions.
Diffusion patterns surprised the researchers. Adoption did not stay inside engineering. Legal, finance, and recruiting teams crossed over in April 2026, and non-engineering departments soon accounted for more than 85% of Codex tokens. OpenAI lists Product, Marketing, and Ops among the primary knowledge-work functions using agents, and notes that over a quarter of the work business users complete with Codex is engineering-shaped: automation, data transformation, tool assembly, debugging, structured analysis. The boundary between “technical” and “business” work is blurring inside the agent layer.
Two structural implications follow. First, the unit of planning shifts upward: instead of briefing a person on a task, you brief an agent on an outcome and a verification rubric. Second, throughput stops being linear with headcount. One orchestrator can run dozens of agents in parallel, which re-asks the classic question of how a team should be sized and structured.
Why marketers should care
For marketing organizations, the implication is less “use AI for copy” and more “redraw the workflow.” Tasks that were always desirable but never staffed — competitive monitoring, creative-performance attribution, batched personalization, cross-platform rule sync — now fit the shape of a delegatable agent task. The marketing function that learns to scope work into agent-readable briefs, with clear inputs and verifiable intermediate artifacts, gets to run many experiments in parallel instead of queuing them through a single spreadsheet.
How to use it
- Hand off research and reporting. Competitive scans, comment clustering, weekly report drafts — anything with structured inputs and a checkable output table is agent-ready today.
- Keep judgment on the human side. Brand positioning, resource negotiation, and the final call on sensitive or compliance-adjacent content remain human decisions; agents compress surrounding execution, they do not make the call.
- Test one parallel workflow. Pick a single channel or audience segment, brief an agent end-to-end, and measure against the human baseline before scaling.
// END OF LOG