If you have been following AI closely, you have probably noticed a subtle but unmistakable shift in how the major players are talking about agents. The demos are over. The infrastructure race has begun.
In the past two weeks alone, OpenAI launched ChatGPT Work, an agent that can run multi-step projects for hours while you are in meetings. Google shipped background execution and remote MCP support for its Gemini managed agents. Mistral moved coding agents to the cloud with remote async agents powered by its new Medium 3.5 model. And NVIDIA rolled out verified agent skills and a CPU purpose-built for agentic workloads.
This is not hype. This is the moment when agentic AI becomes a production technology.
OpenAI: From Codex to ChatGPT Work
When OpenAI released Codex in 2025, it was positioned as a coding agent. Developers loved it. What OpenAI did not expect was how quickly everyone else would adopt it.
In a new research paper, OpenAI reveals that Codex usage among non-developers has exploded. Individual non-developer users are up 137x since August 2025. Inside OpenAI itself, the average non-engineering worker now generates over 85% of their output tokens on Codex rather than ChatGPT. Legal, finance, and recruiting departments crossed the majority-Codex threshold around April 2026.
That is the backdrop for ChatGPT Work, which launched on July 9. Unlike Codex, which started as a coding tool and gradually expanded, ChatGPT Work was built as a general-purpose work agent from day one. It can:
- Connect to Slack, Microsoft Teams, Google Drive, SharePoint, CRMs, calendars, and internal tools via plugins
- Run scheduled tasks — for example, reviewing new Slack updates each week, monitoring dashboards every morning, or updating presentations when new customer feedback arrives
- Work across the desktop with a built-in browser, accessing local files and web-based tools in the same session
- Create slides, sheets, docs, and even interactive Sites (public beta) that can be shared via URL
- Operate for hours in the background, breaking complex projects into steps and completing them independently
Perhaps most telling: OpenAI says nearly 100% of teams inside the company, including finance and sales, now use ChatGPT Work and Codex. A sales team used it to turn a discovery call into a tailored proof of concept within 24 hours — a process that normally takes weeks.
Google: Managed Agents Go Production-Ready
On July 7, Google announced a major expansion of its Managed Agents in the Gemini API. The updates directly address the friction points developers have been hitting when trying to ship production agents.
The standout additions:
- Background execution: Pass
background: trueand the agent runs asynchronously on the server, returning an ID you can poll for status. No more holding HTTP connections open for long tasks. - Remote MCP servers: Connect managed agents directly to remote Model Context Protocol servers — your private databases, internal APIs — without writing proxy middleware.
- Custom function calling: Mix built-in sandbox tools with your own client-side functions, with step matching so the right code runs in the right place.
- Credential refresh: Rotate short-lived API keys without losing session state, installed packages, or cloned repositories.
These are exactly the kinds of primitives that make the difference between a prototype and a product. Google is treating agents as asynchronous workers, not chatbots with tool access.
Mistral: Remote Agents in the Cloud
Mistral has been steadily building out its agent stack, and on July 7 it took a leap forward with remote agents in Vibe. Coding agents that used to live on your laptop can now run in the cloud, asynchronously, in parallel.
Powering this is Mistral Medium 3.5, a new 128B dense model with a 256k context window that merges instruction-following, reasoning, and coding into a single set of weights. It scores 77.6% on SWE-Bench Verified and runs on as few as four GPUs for self-hosting. Reasoning effort is now configurable per request — the same model can handle a quick chat reply or a multi-hour agentic run.
Mistral also introduced Work mode in Le Chat, a new agent that handles complex multi-step tasks like research and cross-tool workflows. Sessions can be started from the CLI or directly in chat, and they keep running while you step away. When done, the agent can open a pull request and notify you via Slack.
NVIDIA: Governance and Hardware for the Agent Era
If OpenAI, Google, and Mistral are building the agent interfaces, NVIDIA is building the plumbing.
On July 7, NVIDIA introduced its Vera CPU, explicitly designed to accelerate agentic workloads. Unlike general-purpose processors, Vera is optimized for the specific mix of inference, tool use, code execution, retrieval, and orchestration that agents require. It is part of a broader full-stack push that includes Dynamo for inference optimization and NeMo Guardrails for runtime safety.
Earlier, NVIDIA launched verified agent skills — a governance layer for the capabilities agents use. Each verified skill includes a Skill Card documenting what it does, who built it, its dependencies, known risks, and limitations. Skills are scanned by SkillSpector before publication and signed for authenticity verification. The idea is simple: if an agent is going to run code, access databases, and make API calls on your behalf, you should know exactly what capabilities it has and where they came from.
NVIDIA also released open data products for agent training — over 10 trillion tokens spanning software engineering traces, tool-use failures, multi-step reasoning, safety, and workflow execution. The message is clear: agent behavior needs to be inspectable, and that starts with the data.
OpenAI Builds Its Own Chip
On June 24, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom AI accelerator designed from the ground up for LLM inference. Developed in just nine months (with OpenAI’s own models accelerating the design process), Jalapeño is part of a multi-generation roadmap to deploy gigawatt-scale data centers.
Early testing shows performance per watt substantially better than current state-of-the-art. The chip is designed around the specific memory movement, networking, and serving patterns that frontier models require — not a repurposed GPU, but a blank-slate architecture for modern inference.
The timing matters. As agents consume more tokens, run longer, and operate in parallel, inference efficiency becomes the bottleneck. Custom silicon is how you scale from millions of chat interactions to billions of agent-hours.
What This Means
Three patterns are converging:
- Agents are becoming the default AI interface, not chatbots. OpenAI’s own data shows workers generating 99.8% of their weekly output tokens through Codex. That is a structural shift in how people interact with AI.
- Agent infrastructure is getting real primitives — background execution, credential management, async orchestration, remote tool access, governance layers, and custom silicon. The toy phase is over.
- Non-developers are the fastest-growing agent users. The tools are expanding from coding into finance, sales, recruiting, and general knowledge work. Agentic AI is not a developer tool anymore; it is becoming general-purpose compute.
If you are building with AI, the question is no longer “should I use agents?” It is “which agent infrastructure should I build on?”
The options are multiplying fast. OpenAI has the deepest integration with existing work tools. Google has the most polished managed-agent primitives for developers. Mistral offers open weights and self-hosting. NVIDIA provides the governance and hardware stack underneath everything.
The race is on. And unlike the model-leaderboard battles of 2024 and 2025, this one is about who can make agents actually work — reliably, securely, and at scale — in the systems people already use.