Sam Altman said it plainly. In 2025, AI agents would join the workforce. OpenAI delivered.
Operator arrived in January 2025. The tool takes control of a browser. It clicks. It scrolls. It types. Users on the $200-a-month ChatGPT Pro plan gained early access. TechCrunch reported the launch. Tasks once handled by assistants now run autonomously. Grocery orders. Expense reports. Restaurant bookings.
But the real story runs deeper. OpenAI didn’t stop at one agent. It built platforms. It upgraded models. And it targeted the enterprise heart. Today agents don’t just chat. They execute. They plan. They integrate into sales teams, engineering groups and finance departments. The shift happened faster than many predicted.
From Chatbots to Autonomous Actors
ChatGPT changed how people write and research. Agents change how work gets done. Operator uses its own virtual browser. It interacts with websites the way humans do. No special APIs required for many sites. That flexibility carries risks. And limits.
OpenAI upgraded the underlying model months later. It replaced the GPT-4o version with one based on o3. Performance improved. The New York Times tested it. The system proved brittle at times. Erratic in spots. Yet the direction stood clear. The publication noted it pointed toward powerful future agents.
By July 2025 OpenAI expanded the concept. A new ChatGPT agent handled spreadsheets. It built PowerPoints. It automated presentations and analyses. The Wall Street Journal described how this put OpenAI in closer competition with Microsoft. The pace accelerated.
Enterprises watched closely. They wanted control. They needed governance. OpenAI responded in early 2026. It launched Frontier. The platform lets companies build and manage agents. Treat them like human employees. Track performance. Assign tasks. Integrate with existing systems. TechCrunch detailed the end-to-end offering. It works with agents built outside OpenAI too. An open approach.
Codex grew fast. OpenAI reported it reached 3 million weekly active users by April 2026. APIs process more than 15 billion tokens per minute. GPT-5.4 powers record engagement in agentic workflows. Customers include Goldman Sachs, Phillips and State Farm. Existing users such as DoorDash, Thermo Fisher and LY Corporation expanded use. OpenAI’s own announcement laid out the numbers.
Sales teams deploy agents that research prospects. Score them. Send tailored emails. Update the CRM. Engineering groups run multi-agent systems that handle work end to end. One agent plans. Another codes. A third tests. The coordination feels closer to a team than a single tool.
Yet challenges remain. Reliability. Security. Cost. Agents still fail on complex, long-horizon tasks. They need guardrails. Enterprises worry about data leaks when agents browse the open web. OpenAI added sandboxes. Monitoring tools. The April 2026 Agents SDK update brought safety features and evaluation harnesses. TechCrunch covered those enhancements.
Payment integration arrived this month. Visa partnered with OpenAI. Agents now hold tokenized credentials. They make secure purchases within limits set by users. The network identifies the agent. Banks authorize in real time. Simon Taylor, fintech analyst, called it a potential mainstream moment for agentic commerce. Recent X discussions highlight the development.
By late 2025 agents had moved from experiment to infrastructure. The Conversation reviewed the year. Models from OpenAI, Anthropic, Google and Chinese labs improved reasoning. Tools turned browsers into active participants. Perplexity, the Browser Company and others followed similar paths. That analysis captured the momentum.
OpenAI leads but doesn’t stand alone. Microsoft pushed its own agent service at Build 2025. It emphasized multi-agent orchestration and governance. Anthropic and others invest heavily. Competition drives faster releases. Better memory. Stronger planning.
Usage data tells part of the story. One service saw agents used for code writing 17 percent of the time. Research took 10 percent. Image creation, document generation and brainstorming followed. The New York Times reported these patterns in June 2026. Its article asked what agents actually do. The answer? More real work than before.
Inside OpenAI the transformation runs deep. Employees use Codex across functions. Product leaders describe it as a software engineering teammate. It participates in the full development lifecycle. Growth exploded 20 times in recent months. Trillions of tokens processed weekly.
The enterprise bet looks clear. Companies won’t rip out existing systems. They will layer agents on top. Manage them centrally. Measure output like any other worker. Frontier aims to make that possible. So does the expanded Agents SDK.
But adoption brings new questions. Who owns the work an agent produces? How do you audit decisions made across multiple steps? What happens when an agent books the wrong flight or approves the wrong invoice? Legal and compliance teams scramble to catch up.
OpenAI continues to ship. Memory improvements for ChatGPT arrived recently. New capabilities in specialized models. Codex plugins for sales, analytics, design and investment banking. Annotations let users edit specific sections of agent output. Sites turn plans into interactive applications.
The technology no longer feels futuristic. It sits in workflows today. Sales reps close deals with less manual research. Developers ship code faster. Analysts produce reports that once took days. The productivity gains appear real even if uneven.
Critics point to overhype. Many early agent demos still require heavy human oversight. Hallucinations persist in subtle forms. Integration with legacy enterprise software remains messy. Yet the trajectory holds. Each model upgrade brings longer reliable task chains. Better tool use. Fewer mistakes.
Visa’s move signals broader acceptance. Banks and payment networks see agents as future customers. Not threats. They build identity and trust layers so agents can act with permission. That infrastructure matters as much as the models themselves.
OpenAI’s advantage lies in distribution. Hundreds of millions already use ChatGPT. Agents appear inside the same interface. No new login. No new vendor. That ease accelerates adoption. Enterprises already paying for enterprise licenses can test agents with minimal friction.
The coming months will test maturity. Can agents handle entire processes with minimal intervention? Will enterprises trust them with sensitive data? How quickly do competitors match the capabilities?
One thing looks certain. The age of agents isn’t approaching. It has arrived. Companies that learn to manage digital workers alongside human ones will hold the edge. OpenAI supplied the tools. Now the hard part begins. Making them work at scale. Reliably. Profitably.
WebProNews is an iEntry Publication