The agentic AI era is no longer a preview. Over the past month, the boundary between “AI assistant” and “autonomous coworker” has dissolved at an accelerating pace. From OpenAI’s own workforce converting almost entirely to agentic workflows, to Mistral unifying work and code under a single agent, to custom inference silicon purpose-built for LLMs, the infrastructure and behavior of AI agents is shifting from experiment to default.
Here is what changed in Agentic AI in late June and early July 2026.
OpenAI’s Workforce Is Now Majority Agentic
In a detailed blog post published on June 25, OpenAI revealed internal usage data that should settle any debate about whether agents are a real productivity tool or a novelty. By May 2026, 80.6% of individual Codex users had made at least one request estimated to exceed 30 minutes of human work. 70.2% had made one exceeding an hour. And 25.6% had delegated a task estimated at over eight hours.
The pattern inside OpenAI itself is even more striking. Codex now accounts for 99.8% of weekly output tokens generated within the company. The average engineer generates 99% of their output tokens through Codex rather than ChatGPT. Legal, Finance, and Recruiting crossed over to majority Codex usage around April 2026, and the average non-technical worker now generates more than 85% of their tokens through the agent.
Most notably, non-developer adoption has outpaced developer adoption. Since August 2025, non-developer individual users rose 137x, organizational non-developers 189x, and internal non-developers 12x. Workers in business functions are now regularly using Codex for engineering and coding tasks—over one-fourth of work done by business-function workers with Codex was technical.
The implication is clear: agents are not replacing a single job function. They are dissolving the boundaries between them.
GPT-5.6 Sol: Reasoning, Subagents, and a Phased Release
On June 26, OpenAI previewed the GPT-5.6 series: Sol (flagship), Terra (balanced, 2x cheaper than GPT-5.5), and Luna (fast and affordable). The headline feature is ultra mode, which goes beyond single-agent reasoning by leveraging subagents to accelerate complex work.
Sol sets new state-of-the-art results on Terminal-Bench 2.1 (command-line workflows requiring planning and tool coordination), GeneBench v1 (long-horizon genomics), and ExploitBench (cybersecurity vulnerability research). On ExploitGym, a benchmark created by UC Berkeley researchers in collaboration with OpenAI, GPT-5.6 Sol demonstrated strong improvements while using approximately one-third the output tokens of Mythos Preview.
OpenAI paired the capability jump with its most robust safeguard stack to date. Real-time cyber and biology misuse classifiers evaluate output as it is generated. Higher-risk cases trigger pauses while a larger reasoning model reviews the conversation. The model is trained to refuse prohibited requests, including disguised intent and jailbreak attempts. OpenAI also previewed the model with the U.S. government before launch, though the company explicitly stated it does not believe this kind of government access process should become the long-term default.
Jalapeño: OpenAI’s Custom Inference Chip
On June 24, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom Intelligence Processor—a chip designed from scratch for LLM inference. Developed from design to manufacturing tape-out in just nine months, Jalapeño is a blank-slate accelerator optimized for the kernels, memory movement, networking, and serving patterns that matter for frontier AI models.
Early testing shows performance per watt substantially better than current state-of-the-art accelerators. The architecture reduces data movement and balances compute, memory, and networking resources to achieve realized utilization closer to theoretical peak performance. Broadcom’s silicon implementation and Tomahawk networking silicon are key enablers for large-scale production.
The significance extends beyond speed. By designing the inference stack—from chip architecture through kernels, scheduling, and product experience—OpenAI is tightening a flywheel where better infrastructure drives compute efficiency, which enables better models, which drives more usage and revenue, which funds the next generation of infrastructure. Jalapeño is the first step in a multi-generation compute platform planned for gigawatt-scale deployment with data center partners beginning in 2026.
Mistral Vibe: One Agent for Work and Code
While OpenAI is doubling down on Codex, Mistral AI is pursuing a parallel vision with Vibe—a unified agent that handles both long-horizon knowledge work and coding workflows under one interface. Formerly Le Chat, Vibe now offers two modes:
- Work Mode runs multi-step tasks across enterprise tools—catching up on inboxes, running research, drafting deliverables, analyzing structured data, and scheduling recurring processes.
- Code Mode handles remote coding sessions from feature development to merged pull requests, running in isolated sandboxes that persist while your machine is off.
Vibe integrates with Google Workspace, Outlook, SharePoint, Slack, GitHub, and custom connectors. It maps out plans before execution, streams progress visibly, and supports reusable skills via open standards. A new VS Code extension brings the coding agent directly into the editor, reading and editing files across the entire project.
The underlying models—optimized for reasoning, agentic tasks, tool calls, and coding—reflect Mistral’s bet that the future of work is not separate tools for separate tasks, but one agent fluent in your context.
Anthropic, Glasswing, and Industry-Wide Safety
Anthropic’s June 30 announcement brought two developments. Fable 5 returns globally on July 1, continuing the company’s model release cadence. More structurally, Anthropic proposed an industry-wide framework for scoring jailbreak severity, developed in partnership with Amazon, Microsoft, Google, and other Glasswing partners.
This is notable not just for the technical standard, but for the coordination signal. As agentic models gain capability—handling longer tasks, accessing more tools, and operating with less human supervision—jailbreak severity becomes a systemic risk question, not a model-by-model concern. A unified scoring framework is a prerequisite for meaningful cross-platform safety auditing.
OpenClaw Skill Workshop: Making Agent Work Reusable
Not every agentic advance comes from a frontier lab. On June 3, OpenClaw introduced Skill Workshop, a system that turns agent work into reusable, reviewable skills. When an agent creates or revises a skill, it starts as a pending proposal—not an immediate write to live behavior.
The workflow is deliberate: the agent drafts a skill, the user reviews it in a board or day view, tweaks wording or adds steps, and then applies or rejects it. Support files—templates, scripts, examples—travel with the proposal and are scanned before activation. This mirrors how human teams handle process changes, and it addresses a real risk: a bad skill is worse than a bad answer because it becomes part of how future work is done.
Skill Workshop also supports revision conversations, so skills can evolve without losing history. The same flow works from chat, the Control UI, channels, and Gateway. For anyone building agentic systems, this is a useful pattern: agent-generated process changes should be proposals first, not writes.
Agentic AI in the Wild
Other signals from the ecosystem reinforce the direction:
- Hugging Face continues to expand its evaluation infrastructure, with new community eval leaderboards and vLLM server deployment on HF Jobs.
- LlamaIndex launched ParseBench, the first document-parsing benchmark designed for AI agents, evaluating parsers across tables, charts, content faithfulness, and visual grounding.
- Braintrust published deep dives on evaluating stateful agents, multi-turn conversations, and agent cost-efficiency—reflecting the industry’s shift from model benchmarking to agent benchmarking.
- OpenClaw’s Skill Workshop and NVIDIA partnership for skill security scanning show the tooling layer maturing alongside the model layer.
What This Means
The agentic AI story of mid-2026 is not about one breakthrough. It is about convergence: models getting better at long-horizon tasks, infrastructure getting cheaper and more efficient, interfaces getting unified, and safety frameworks getting industry-wide. The result is that agents are moving from “sometimes useful” to “primary tool for work”—and not just for developers.
The OpenAI data is the strongest empirical evidence yet that this shift is real and accelerating. When a company’s own workforce delegates 99.8% of its AI usage to agents, the chatbot era is officially ending. What replaces it is a world where AI does the work, humans supervise the agent, and the boundary between job functions becomes as fluid as the models can make it.