Sam Altman has long talked about AI systems that do more than answer questions. They act. In January 2025, OpenAI delivered on part of that vision with Operator, a research preview of an agent that uses its own browser to complete tasks for users. It could visit websites, click buttons, fill forms and handle simple workflows without constant supervision.
But that was just the start. Over the past 16 months, OpenAI has rolled out a series of products that move these agents deeper into business operations. Workspace agents arrived in ChatGPT for Business, Enterprise and education users this April. They connect to tools like Slack and Gmail. They gather context, follow established workflows, ask for approvals when needed and get better with use. OpenAI positioned them as the successor to custom GPTs, designed for repeatable work that once required manual repetition and data copying across applications.
Then came Frontier. Launched in February 2026, the platform gives enterprises the ability to build, deploy and manage AI agents that operate with shared context across company systems. Agents gain onboarding, clear permissions and boundaries. They learn from experience. OpenAI calls them AI coworkers. The company even offers services from its own engineers to help large organizations integrate them.
These moves reflect a clear shift. Early chatbots delivered information. Agents now execute. They book travel, process expense reports, research sales leads, update CRMs and coordinate across teams. One OpenAI sales team example involved an agent that scores inbound prospects, sends personalized emails and logs activity in the CRM. No constant prompting required.
Yet the technology remains imperfect. Operator, now integrated into ChatGPT as the ChatGPT agent, still shows brittleness. It can misclick, get stuck in login flows or fail on complex, multi-step processes. Early testers noted its tendency to generate unnecessary code or repeat steps before simplifying. Similar issues appear across the industry. But the pace of improvement is rapid.
Competitors have responded in kind. Google released new agent-building tools in April 2026 to automate tasks and challenge OpenAI and Anthropic directly. The Mercury News reported the effort as part of a broader race for enterprise dominance. Anthropic updated its own computer-use capabilities. Chinese developers push open-weight models that power agents at lower cost.
OpenAI itself updated its Agents SDK in mid-April. The changes added harness and sandbox features aimed at safer, more reliable agents for business. TechCrunch described the enhancements as critical for companies seeking to move agents into production without exposing themselves to prompt injection or unintended actions.
Adoption data tells its own story. By early 2026, Codex, OpenAI’s coding agent, reached 3 million weekly active users. APIs processed more than 15 billion tokens per minute. GPT-5.5, released around the same period, improved reasoning and speed for enterprise and scientific work. The company framed it as progress toward a unified super app that blends chat, coding, search and agent capabilities. MarketingProfs noted the release alongside workspace agents as evidence of demand for systems that perform work rather than merely assist.
Surveys show organizations no longer debate whether to build agents. They focus on how to deploy them at scale with reliability. The LangChain State of AI Agents report found that most teams use multiple models. OpenAI’s GPT series leads, but Claude, Gemini and open-source options fill specific roles based on cost, latency and task type. Fine-tuning remains rare. Research and deep-research agents rank among the most common patterns.
Enterprise interest has grown fast. Customers including Goldman Sachs, Phillips, State Farm, GitHub, DoorDash and Thermo Fisher now run multi-agent systems. Some handle end-to-end engineering tasks. Others manage sales, marketing or operations workflows. The next phase of enterprise AI, according to OpenAI, centers on agents that integrate with systems of record, maintain institutional memory and operate with governance. OpenAI highlighted this shift in an April 2026 update.
Risks remain front and center. Agents that control browsers or connect to internal tools can make mistakes at scale. They might leak data, take unauthorized actions or amplify biases in training data. Sandboxing helps, but full autonomy in sensitive domains still makes compliance teams nervous. OpenAI has stressed iterative safety work, including better observation, permission systems and human-in-the-loop checkpoints.
Cost presents another hurdle. While workspace agents started free until early May 2026, credit-based pricing followed. Enterprises running fleets of agents across thousands of tasks face real expenses. Those costs compound when agents generate excess tokens through trial-and-error loops before completing a job.
But the upside appears substantial. Companies report agents completing weeks of human work in days. Marketing teams automate reporting and outreach. Developers let agents handle routine coding before human review. Operations groups coordinate across Slack, email, CRMs and project tools without constant handoffs. The pattern repeats across functions.
Analysts expect further convergence. Browser-based agents like Operator will likely merge more tightly with workflow-specific agents in platforms like Frontier and workspace tools. Multi-agent systems, where specialized agents hand off subtasks to each other, are already in use at some early adopters. Infrastructure providers race to supply the compute. NVIDIA’s delivery of custom Vera CPUs to OpenAI, Anthropic and others signals the hardware buildout required for reliable agent performance at scale.
Recent conversations on X highlight both excitement and skepticism. Developers debate runtime control layers that current frameworks lack. Some note agents still produce bloated code that requires cleanup. Others point to practical wins in research, data gathering and simple transaction flows. One user demonstrated Operator searching Google Maps reviews for stores with distinctive writing styles, watching it navigate, read and avoid login traps with mixed success. The video captured both capability and current limits in one clip.
The original Financial Times reporting on OpenAI’s agent ambitions, published before many of these launches, captured the company’s early conviction that such systems would reach mainstream use quickly. Financial Times outlined timelines and competitive pressures that have largely played out as described.
Where this leads depends on execution. If OpenAI, Google, Anthropic and others can tame reliability, cost and safety issues, agents could reshape job descriptions across knowledge work. Routine coordination tasks may migrate to software. Human roles could tilt toward oversight, exception handling and strategic direction. That transition carries economic and social questions that extend far beyond technology choices.
For now, the products keep arriving. Frontier expands. Workspace agents gain integrations. SDK improvements roll out. Each iteration brings agents closer to the independent workers their creators envision. The experiments have ended. Implementation has begun.
WebProNews is an iEntry Publication