Field note 05
The Shift From Chat to Agentic Work
New usage research suggests the important change is not how often people talk to an agent, but how much complete work they are willing to entrust to it.
The earliest AI coding workflow looked like a faster search box. Ask a question, get a snippet, paste it into an editor, and continue working. The current workflow is beginning to look less like chat and more like delegation.
That does not mean the human disappears. It means the unit of work gets larger.
What is confirmed
On June 25, OpenAI published research on how agents are changing work inside the company. The company reports that every department uses Codex as its primary AI tool. More than 70 percent of users asked Codex in May 2026 to complete at least one task estimated to take a person more than an hour. At the high end of the distribution, some users ran more than 60 hours of Codex turns in a day by working with agents in parallel.
The accompanying paper reports that active Codex users grew more than fivefold during the first half of 2026. More than 10 percent of weekly users managed three or more concurrent agents at least once in a week, and 26.6 percent used skills. The researchers also found that tasks were becoming more complex and longer in horizon.
These are OpenAI's own workplace and product data. They are useful evidence, but they should not be treated as a neutral survey of every industry or team. OpenAI is unusually close to the tools, unusually motivated to use them, and able to change its own product around observed workflows.
Delegation changes the shape of the day
When a task takes five minutes, the natural rhythm is conversation. When it takes an hour, the important skills become specification, checkpointing, and review.
A larger task needs:
- A clear deliverable.
- Access to the right repository and tools.
- Boundaries around destructive or external actions.
- A proof plan that matches the risk.
- A place to record decisions and unfinished questions.
Parallel agents add another requirement: the work must be separable. Giving three agents permission to edit the same central file is not parallelism. It is a merge conflict with a progress bar.
Good parallel work has clean edges. One agent can research source material while another writes tests. One can inspect database migrations while another audits the interface. Each task should produce an artifact the main effort can evaluate.
My take
The dramatic number in the research is 60 hours of agent turns in one human day. The more consequential number may be 26.6 percent using skills. Parallel compute creates volume. Reusable instructions create organizational memory.
A skill can capture a release checklist, a security review, a 3D asset validation pass, or the exact way a company publishes documentation. That turns a good one-time conversation into a repeatable workflow. It also makes the workflow inspectable. A teammate can read the instructions, challenge them, and improve them.
The central management problem is no longer “What should I ask?” It is “What work can be safely delegated, what proof do I need back, and which decisions remain mine?”
That is a healthier question than either blind trust or blanket skepticism.
A rule worth copying
Delegate outcomes, not vague effort. Every agent task must return an artifact, verification evidence, and any unresolved decisions.
Sources and further reading
- OpenAI's report on how agents are transforming work
- The accompanying research paper
- Official Codex use cases
- How to build reusable Codex skills