Agents as Ongoing Machine Labor

This shifts OpenAI from selling access to selling ongoing machine labor. Once Codex and ChatGPT Work move from one prompt and one answer into scheduled runs, multi tool loops, and parallel agents, revenue can rise because the same employee now consumes far more tokens across the day and overnight, even if the company buys no additional seats.1 2 3

• In practice, this looks less like chat and more like software or ops work being delegated. One Codex power user had around twenty daily automations spanning email, Slack, calendar, call review, competitor scans, and release notes, with most outputs reviewed inside Codex before being pushed into other tools.1
• The same pattern appears in teams. At Scale AI, Claude Cowork and Codex were used for scheduled social posting, QC review, dashboards, and computer use evals. The bottleneck was no longer whether the model could answer, but whether multi step chains across several tools stayed reliable enough to keep running.2
• This matters economically because OpenAI already combines seat pricing with usage pricing. Business plans include bundled usage plus extra credits, and the API charges by tokens, so deeper automation raises spend per seat and per workflow. By June 2026, 99th percentile Codex users were generating more than 60 hours of agent work per day across parallel agents.3

The next phase is agents becoming a standing layer inside white collar software, not a chat window people open occasionally. As reliability improves and more work runs on OpenAI managed computers or through the Agents API, token growth should increasingly track tasks completed and processes automated, not just headcount or message volume.2 3