OpenAI says Codex now has more than 5 million weekly users. The more important detail: non-developers make up about 20% of Codex users and are growing more than 3x faster than developers. OpenAI is also adding role-specific plugins, Sites, and annotations for workflows across data analytics, creative production, sales, product design, investing, and banking.

That changes how Codex should be understood. This is no longer only a better coding tool. Codex is being rebuilt around work artifacts: plugins bring in role-specific context, Sites turn outputs into shareable review surfaces, and annotations make feedback part of the artifact itself. For operators, the signal is clear: Codex is moving closer to how teams actually create, review, and improve work.

At Build 2026, Microsoft rolled out a broad agent stack: Microsoft IQ, Work IQ, Web IQ, Scout, MAI-Thinking-1, Agent 365, ASSERT, ACS, MXC, hosted agents, and Surface RTX Spark Dev Box. Mustafa Suleyman also framed MAI around Frontier Tuning, company-specific agents, model control, Microsoft silicon, and cost/performance.

Microsoft is trying to turn agents into enterprise infrastructure that can be controlled: models tied to company context, behavior constrained by security policy, runtime controlled through checkpoints, and cost/performance pulled into the same stack. This is the direction agents need to move if they are going to leave demos and enter production workflows.

Anthropic is expanding Project Glasswing to around 150 new organizations across more than 15 countries, including critical infrastructure sectors such as power, water, healthcare, communications, and hardware.

Anthropic is moving frontier cyber defense into controlled access lanes. Frontier AI is useful enough to be applied to defense for critical infrastructure, but sensitive enough to require trusted deployment, partner access, and governance. That is the other side of agent adoption: the closer AI gets to real work, the more real constraints it needs.

Codex is becoming a clickable enterprise work surface: Simon Smith argues that Codex plugins, annotations, and Sites make enterprise workflows more interactive: plugins can add guided actions, annotations make feedback more precise, and Sites make internal tools easier to build, share, and secure. Treat this as an operator signal: Codex is not only creating outputs; it is turning outputs into things teams can review and edit directly.

OpenAI-backed Opal is building an AI audio gadget: WIRED reports that Opal is rebranding as Opal Electronics, expanding beyond webcams, and building an AI-powered audio product after a $40 million Series B from OpenAI. The signal is that OpenAI's hardware strategy may not revolve only around a single first-party Jony Ive device; it may also include adjacent bets on consumer AI surfaces.

Codex Sites could pressure app-layer startups: Jason Calacanis argues that Codex/Sites threatens Cursor, Lovable, startups, and application-layer partners because infrastructure companies eventually move up the stack. This is a claim-by-source, not a proven market fact, but it captures the platform question: if Codex Sites lets every user create deployed private workspace apps, app-layer AI builders will need a clearer moat.

Codex Mobile is becoming a remote agent control plane: Vaibhav Srivastav says Codex Mobile for iOS added Face ID/passcode lock, Queue or Steer defaults, line wrapping for diffs, SSH connection to Windows machines, /side prompts, and several fixes around follow-ups, reconnecting, archiving, tasks, and host connections. It is a small product signal, but it points to Codex Mobile becoming a control plane for background and remote agent work.

Perplexity is moving Computer toward hybrid local/cloud inference: Perplexity says hybrid agentic inference is coming to Perplexity Computer, letting tasks split between a local model on the user's machine and frontier models in the cloud. This is an important architecture signal: private or context-heavy work can stay local, while harder reasoning routes to the cloud for better cost, privacy, and capability tradeoffs.

Agent 365 wants to make local agents visible to enterprise IT: Microsoft says Agent 365 can surface unmanaged local agents discovered by Defender, Entra, and Intune, including coding agents, AI desktop apps, and local or remote MCP servers. Enterprises are starting to manage agents like assets and identities, not just apps.

OpenAI token demand is becoming a customer cost problem: Axios, citing Sam Altman, reports that OpenAI's top token user consumes about 100 billion tokens per month, while cost concerns are becoming a major customer issue. As AI shifts into proactive and background work, token usage becomes a governance problem rather than a hidden product cost.

Google adds fake-call detection for AI impersonation scams: Google is rolling out Android fake-call detection inside Phone by Google to fight AI deepfake impersonation scams. AI abuse is now mainstream enough that verification is becoming part of default phone UX, not a niche security feature.

If Codex is being built around role plugins, Sites, and annotations, the better workflow is not "prompt once and accept the output."

The better workflow is to turn a task into a reviewable artifact: a Site, dashboard, deck, working doc, or small internal tool. In 30 minutes, you can set up a simple workflow that makes Codex act more like a project workspace.

STEP:

WAY TO EVALUATE:

Last Tuesday, marketing asked Viktor to write the weekly campaign recap, pull performance from Google Ads and Meta, and format it as a PDF for the exec team. Done in four minutes.

That same afternoon, engineering asked Viktor to review three open pull requests on GitHub, cross-reference with the Linear sprint board, and flag anything blocking the release. Posted to private channel before standup.

At 9pm, ops asked Viktor to draft a vendor contract summary from three Notion docs and send it to the team. It was in #ops by morning.

None of them knew the others were using it.

Same colleague. Three departments. That's what changes when your AI coworker lives in Slack, where your whole company already works. It's not a tool one person logs into. It's a teammate everyone messages.

5,700+ teams. SOC 2 certified. Your data never trains models.

"Viktor is now an integral team member, and after weeks of use we still feel we haven't uncovered the full potential." - Patrick O'Doherty, Director, Yarra Web

Start free. $100 in credits →

The Information says OpenAI built Codex into one of its fastest-growing products after falling behind Anthropic in coding, and is now making Codex central to ChatGPT's future. Worth reading because it adds strategic context to today's lead: coding and work execution may be where ChatGPT becomes a durable product surface, not just a general assistant.

Microsoft Foundry lays out why agent trust needs more than benchmarks: teams need policy evaluation, runtime controls, and production observability. Worth reading because it gives a concrete framework for moving agents from demo to governed deployment.

Axios frames Altman's token-user comments as evidence that token demand, compute constraints, and cost are becoming major enterprise adoption issues. Worth reading because proactive agents will make usage governance more urgent, especially when background work scales quietly.

The Verge rounds up the Build 2026 AI announcements across Scout, MAI-Thinking-1, Project Solara, MXC guardrails, Surface RTX Spark Dev Box, and developer updates. Worth reading if you want the broad map of Microsoft's AI push without reading every product post separately.