AI Work Needs A Shared Floor

The strange thing about the agentic workplace is that it may arrive less like a revolution than like uneven office weather.

One person has a coding agent running tests in the background, another has a carefully tuned set of instructions for research and reporting, another is still using the assistant as a more polite autocomplete, and someone two desks over is quietly pretending they know what everyone means by "delegate it to the agent." This is not resistance so much as ecology. New tools rarely enter a workplace as one clean layer. They seep into habits, shortcuts, private experiments, team rituals, and small acts of professional camouflage.

That is what makes OpenAI's new study of Codex usage more interesting than a simple adoption chart. The researchers describe a shift from assistive coding toward more agentic patterns, with users asking Codex to complete larger units of work, operate across repositories, and handle parts of the development loop that used to require more direct steering. Inside OpenAI, usage is much deeper. Workers there use Codex more frequently, for more varied tasks, and with more developed habits around instructions, reviews, and delegated work.

That difference is the part worth sitting with. It is tempting to read it as the usual early-adopter story: the people closest to the tool use it best, then everyone else catches up. But agentic work has a nastier wrinkle. The gap is not only between people who use the tool and people who do not. It is between people who understand how work has been reshaped around the tool and people who inherit the output without sharing the surrounding practice.

That gap can make a workplace feel quietly bilingual. The agent-heavy worker talks in plans, diffs, prompts, context windows, evals, sandboxes, screenshots, retries, and traces. The agent-light worker receives a finished artifact, a pull request, a spreadsheet, a deck, a summary, or a decision memo that has been through a process they cannot quite inspect. Everyone is technically working on the same thing. They may no longer be working inside the same model of how the thing came to exist.

This is where the usual AI productivity story starts to wobble. A person who can move faster with agents may genuinely be doing more useful work. They may also be generating a new kind of coordination debt. Someone has to know which parts were checked, which parts were guessed, which stale assumption survived the run, which convenient abstraction hid a mess, and which decision came from a human with judgment rather than a tool with confidence and no weekend plans. If that knowledge lives only in the head of the power user, the team has not gained a capability so much as relocated part of the work into a private cockpit.

OpenAI's own product direction points at the same problem from the interface side. Codex is being framed less as a single coding assistant and more as a work surface for roles, tools, and workflows, with role-specific plugins, annotations, and shared outputs that let other people inspect and reuse agent work. Microsoft has been making a similar organizational argument in its Work Trend Index: the frontier firm is not merely a company with more AI licenses, but one that redesigns work, management, and human agency around mixed human-agent teams.

The phrase "mixed human-agent teams" sounds tidy, which is how you know it is hiding something damp in the walls. Real teams are not diagrams. They are people with calendars, grudges, half-remembered migrations, onboarding docs nobody trusts, and a Slack channel where the important context arrived as a joke. When a few people begin working through agents and others keep working through older interfaces, the risk is a slow loss of shared feel.

Shared feel is the unglamorous material of good work. It is the team's sense of what counts as done, which shortcut is acceptable, how much evidence a claim needs, which tests matter, which customer complaint is the real one, and when a clean-looking result smells faintly of wishful thinking. Some of this can be written down. A lot of it is learned by touching the work together.

Agentic tools can reduce that contact if we let them. The senior person may stop showing the messy middle because the agent produces a polished first pass. The manager may approve a summary instead of following the decision path. The junior person may learn to prompt around a task before learning why the task has its shape. The adjacent team may receive a completed change without seeing the exploration, the false starts, or the tradeoffs that would have made the result trustworthy.

None of this means teams should slow down for the sake of nostalgia. The old shared floor had plenty of splinters. Many workplaces taught people through exhaustion, obscurity, and heroic debugging rituals that should have been retired with the beige printer. Agentic tools can remove real waste. They can make expertise more reachable and help people cross unfamiliar systems.

But a companion is not a curriculum, and a faster worker is not the same as a more legible team. If agentic work is going to become normal, organizations need to design the floor beneath it. That means shared conventions for when agents can act, how their work is labeled, where assumptions are recorded, which artifacts preserve the path rather than only the answer, and how non-power users can review output without learning an entire private religion of prompts. It means demos should include the handoff, not just the magic trick. It means onboarding should teach the work system, not only the shiny assistant.

It also means treating uneven adoption as a design condition rather than a moral failing. Some people will be faster. Some will be skeptical for good reasons. Some will overdelegate. Some will undertrust. Some will use agents beautifully and then forget to explain what happened because the machine made the quiet part feel quiet. The organization has to keep those differences from becoming a class system of invisible process.

The better future of AI at work is not everyone becoming a tiny solo agency surrounded by tireless software interns. That sounds exciting until you need to maintain anything together. A sturdier future is one where delegation does not erase the trail, speed does not privatize judgment, and people can still learn the texture of the work from one another.

Agentic work needs a shared floor because work is never only output. It is memory, trust, standards, apprenticeship, repair, and the ordinary dignity of understanding what you are being asked to stand behind. If AI changes who touches the work, it has to change how the work remains touchable.