Codex has to survive the office floor now

The Samsung announcement is fresh, dated June 21. Under the agreement, ChatGPT Enterprise and Codex go to every Samsung Electronics employee in Korea and every DX employee worldwide. OpenAI also says the tools will sit inside Samsung's security policies and governance framework through enterprise controls for data protection, users, access, and security.

The scale matters, but the job list matters more. OpenAI names R&D, manufacturing, marketing, product development, corporate functions, and software work. Those are very different rooms. The same agent has to help a developer review code, a marketer shape campaign material, and an operations team turn an ugly recurring task into something repeatable.

OpenAI also says more than 5 million people now use Codex weekly, and that Codex weekly active users in Korea have grown nearly 800% since February 1. Treat those as vendor-reported numbers, but they point at the same direction: Codex is being pushed into normal work, not kept in the developer corner.

OpenAI's June 2 Codex update already set this up. The company said non-developers are about 20% of Codex users and are growing more than three times as fast as developers. It introduced role-specific plugins, annotations, and Sites, a preview feature for creating hosted interactive websites and apps that can be shared inside a workspace.

That is a founder-message shift hiding in a product note. The buyer is asking whether a team can point an agent at a spreadsheet, a weekly review, a sales process, or a messy product question and get a small internal app back. Something other people can open. Something with a URL. Something that survives comments from a boss who does not care how impressive the model looked while making it.

Annotations are also a quiet tell. First drafts are cheap now. The real work is the second pass: fix this chart, change this section, use the real column name, keep the part we liked. Agent products that make refinement precise will feel less like chat and more like tools.

OpenAI is not alone here. Google launched Managed Agents in the Gemini API in May, letting developers spin up an Antigravity agent in an isolated Linux environment with one API call. The agent can reason, use tools, execute code, browse the web, and keep session state for follow-up calls.

Google's interesting choice is how it lets teams define agents. Instead of forcing every builder to write orchestration code from scratch, it supports versionable files like AGENTS.md and SKILL.md. That is a boring-looking feature with a sharp consequence: agent behavior starts to look like product configuration. You can review it, diff it, ship it, and break it.

The big vendors are turning the same pieces into platform primitives: sandbox, memory, tool access, role instructions, app output, and sharing. Once those pieces are rented, the smaller builder's advantage has to move up the stack. Taste, workflow focus, receipts, and distribution matter more.

There is a useful brake in Anthropic's May survey of 1,260 quantitative social scientists. Eighty-one percent of respondents had tried generative AI in research, but only 20% regularly used coding agents more than once a week. Coding-agent users reported more project starts and working papers, yet Anthropic found no evidence that they were submitting more journal papers or resubmitting faster. The authors also warn that the productivity comparison is descriptive, not causal.

That is probably closer to the workplace reality than the launch posts. People will try chat quickly. They will adopt agents more slowly, because agents ask for a different kind of trust. They need access to files, tools, data, browser sessions, payment systems, or internal repos. They leave artifacts behind. Sometimes those artifacts become someone else's problem.

So the Samsung story should not be read as agents won the office. It means the office is now a deployment surface. The next fight is whether normal teams can tell what an agent did, who owns the thing it made, and how to shut it down when it quietly becomes important.

If you are building around agents, stop pitching the agent as the unit. Pitch the smallest work loop someone already hates. Weekly sales cleanup. Product feedback triage. A manufacturing checklist that becomes a shared tracker. A research note that turns into a reviewable dashboard. The magic is not that the agent can code. The magic is that the person did not have to become the software team for a small internal need.

Then add the receipt before the demo gets too shiny. Show the input sources, permissions touched, tool calls that mattered, files or apps created, current owner, review status, and the cost of the run. If the agent created a hosted app, the artifact should say who can open it and what will break if the source data moves.

The boring enterprise features are not a tax on the fun part. They are what let the fun part survive contact with a real company.

Mina Torres likes the normal-user opening, but only if the product drops the platform language. Her version of the pitch is simple: press a button, get the tracker or page you needed, see what changed, and undo it without calling the AI person.

Priya Rao wants the rollout measured after review, not at generation time. Her dashboard would track request-to-usable-artifact rate, review edits, abandoned runs, security escalations, duplicate tools created, and whether teams keep using the app two weeks later.

Cass Bell is watching for the junk-drawer failure mode. If every department can spawn little tools inside an agent canvas, someone has to delete the dead ones, patch the risky ones, and explain why three dashboards now disagree.