On June 2, OpenAI held an event called “Intelligence at Work.”

For a long time, Codex was easy to explain. It was OpenAI’s coding agent. You gave it a repo, asked it to fix a bug, write a feature, explain a file, or create a pull request. But the latest Codex update shows a bigger shift.

OpenAI has added three useful things: role-based plugins, Codex Sites, and annotations. On paper, these sound like product features. In reality, they point to something bigger.

Codex is slowly moving from a coding tool to a work tool.

Here’s the number that caught my attention: non-developers, including analysts, marketers, designers, product managers, investors, and bankers, now make up 20% of Codex’s 5 million weekly users. And they’re growing 3x faster than developers. Data analysis usage among knowledge workers jumped 110% week over week.

Codex was built as a coding assistant. What it’s becoming is something closer to an operating system for knowledge work.

Three updates shipped on June 2. Each one changes something specific. Together, they point somewhere worth understanding.

One Topic: Codex Sites, Plugins, And Annotations

1. Plugins make Codex understand your role

OpenAI launched six new plugins that bundle 62 business apps and 110 automated skills into role-specific packages. No coding required.

The six roles: Data Analytics, Creative Production, Sales, Product Design, Public Equity Investing, and Investment Banking.

Think of each plugin as a pre-built AI workflow for your job. The data analytics plugin connects Snowflake, Tableau, and Databricks. The creative production plugin wires up Figma, Canva, and Shutterstock. The sales plugin talks directly to Salesforce, HubSpot, and Outreach.

What’s interesting about the product design plugin specifically is its depth. It can generate interactive prototypes from a URL or a screenshot, run visual QA across screen sizes, and export directly to Figma with full layer structure. If you’re a product manager or designer, that cuts the back-and-forth with developers significantly.

Five more roles are already on the roadmap: Corporate Finance, Marketing Strategy, Private Equity, Strategy Consulting, and Legal. That list tells you where OpenAI is heading with this.

2. Codex Sites turn AI output into a shared team space

Codex Sites is still in preview, available only on Business and Enterprise plans, but it’s worth paying attention to now.

The idea is simple: describe something you want, and Codex builds a hosted, shareable web app from it. Just a URL you can send to anyone in your workspace.

Think dashboards, scenario planners, project trackers, launch hubs. The kind of things that today live in a fragile Google Sheet or a 40-tab spreadsheet nobody fully understands.

A financial analyst could take a static model, describe what they want, and get back a live web app where a team lead can tweak assumptions in a browser. A product team could get a feedback portal up for a prototype without filing an IT request. Codex Sites deploys to Cloudflare’s edge, so access is fast and the security model is workspace-native only people signed into your ChatGPT workspace can view it.

The practical implication here is what I wrote about in the Hermes Agent setup: private agent intelligence has real value, but it only scales when it becomes shared intelligence. Sites is the bridge between the two. A morning brief your agent compiles becomes a dashboard your whole team opens.

It won’t replace Retool or complex internal tools anytime soon. But for fast, task-specific apps that most teams currently just don’t build because the overhead isn’t worth it, that calculus just changed.

3. Annotations make AI feedback more natural

This one looks small, but it changes the interaction.

Annotations let you highlight a specific part of something and ask Codex to fix only that part. Previously this worked for code and Markdown files. Now it extends to spreadsheets, slides, and documents.

In practice: you highlight a chart in a slide and ask for a clearer label. You select a cell range in a spreadsheet and ask for a revenue visualisation. Codex makes that specific change without touching anything around it.

Before this, making a targeted edit to an AI-generated output often meant regenerating the whole thing and hoping the parts you liked survived. That’s a real friction point, and annotations remove it. The interaction model shifts from “generate and hope” to something closer to how you’d work with a good editor: point, describe, iterate.

The Bigger Pattern and What this really means

Knowledge work has three bottlenecks that haven’t changed much in years. Tool fragmentation – you’re running eight apps that don’t talk. Static deliverables – slides go stale the moment they leave a meeting. All-or-nothing edits – you can’t fix one thing without risking everything else.

Plugins address the first. Sites addresses the second. Annotations address the third.

The question worth sitting with: what’s one thing you build repeatedly, in a tool that makes it harder than it needs to be, that you’d build differently if you could describe it in plain language instead?

That’s where to start.

Interested in travel or photography, read last week’s LensLetter newsletter about lens flare and how it affects your photo?

Read last week’s JustDraft about research agent to get morning brief using Hermes Agent.

Two Quotes to Inspire

Software used to be an asset you maintained. It’s becoming an interface you generate. The teams that understand this earliest will move differently.

A good leader does not remove every problem. They remove the confusion that makes small problems feel bigger than they are.

One Passage Summary From My Bookshelf

Christensen’s central argument is that well-managed companies fail not because they make mistakes, but because they do everything right, they listen to their best customers, invest in their most profitable products, and ignore markets that seem too small or too immature to matter. This is the dilemma. The very practices that make a company successful also make it vulnerable to disruption from below.

What makes this book remarkable is how precisely it identifies the pattern. Disruptive technologies rarely arrive by beating incumbents at their own game. They arrive by serving a segment the incumbent wasn’t paying attention to usually a smaller, less profitable one, and then moving upmarket once the technology matures. By the time the original leader notices, the ground has shifted. This is the book that explains why strong companies lose, and it’s more relevant to what’s happening in enterprise software right now than most recent AI analysis.

From The Innovator’s Dilemma by Clayton M. Christensen