Ok, now the one that everyone is talking about.
OpenAI has launched ChatGPT agent, a new capability that allows the artificial intelligence to perform complex tasks autonomously, utilizing its own virtual computer. Users can now request actions such as analyzing calendars, planning meals, and creating presentations, with the AI able to navigate websites and execute commands efficiently.
This release integrates features from previous models, combining the web interaction abilities of Operator with the analytical strengths of deep research. Notably, ChatGPT agent has achieved significant performance metrics, scoring 41.6 percent on the Humanity’s Last Exam, which evaluates AI against expert-level questions. The model can also outperform previous iterations in various real-world task benchmarks, including a 45.5 percent success rate in editing spreadsheets, which surpasses tools like Microsoft Excel’s Copilot.
However, the company warns of potential risks, including prompt injections that could lead the agent to take unintended actions with private data. OpenAI has implemented safeguards in response to these concerns, especially as the capabilities of the ChatGPT agent have advanced significantly. Currently, this feature is available to select users in the United States, with plans to expand to other regions in the future.
Microsoft’s AI assistant, Copilot, is struggling to gain traction in the market, lagging significantly behind OpenAI’s ChatGPT. Despite being integrated into Windows and the Microsoft 365 ecosystem, Copilot has only accumulated about 79 million downloads, while ChatGPT has surpassed an impressive 900 million downloads, according to data from Bloomberg and Sensor Tower. The situation has been exacerbated by Microsoft’s decision to split Copilot into two versions—work and personal—leading to functionality issues for users transitioning between versions. Furthermore, Copilot lacks access to basic system-level controls on desktop platforms, which has hindered its usability compared to competitors like Apple’s Siri on iOS or Gemini on Android.
And OpenAI isn’t the only Agentic release. The launch of Comet, an AI-powered web browser by Perplexity, marks a significant shift in how users interact with the internet, challenging the dominance of Google Chrome. Comet aims to streamline the browsing experience by allowing users to engage in conversational queries rather than navigating through traditional links, effectively aiming to “collapse complex workflows into fluid conversations.” This new browser operates on the concept of “agentic AI,” performing tasks autonomously and understanding user intent to execute multi-step actions. According to reports, OpenAI is also preparing to unveil its own AI browser.
Why do we care?
OpenAI’s ChatGPT Agent (and Perplexity’s Comet) are the clearest signs yet of AI systems moving from assistants (waiting for instructions) to autonomous actors (taking actions on their own).
This matters because:
- Agents = Workflow collapse. Instead of multiple tools for calendaring, browsing, presentations, etc., the agent becomes a meta-layer across apps. For SMBs, this promises efficiency, but it also disrupts the traditional IT stack.
- Browsers are the new OS. With Comet and the rumored OpenAI browser, the battleground is shifting. If the browser is running autonomous agents with system-like access, it becomes the most privileged application on a device. That’s a massive new attack surface.
- MSPs are on the hook. Clients will soon expect guidance on deploying (or resisting) these tools. Can agents safely manage calendars, CRM data, or even execute API calls? What happens when one gets hijacked via a prompt injection or abused by an insider?
This isn’t just AI hype—it’s a shift in where control and risk sit in the IT environment. (And that’s assuming this all works.)
Microsoft’s struggle highlights the problem. Even with Copilot’s deep OS integration, users are frustrated. If OpenAI agents are bolted onto systems without deep integration, they may face the same functional limits—and frustrated clients may come to their MSPs to “fix” what isn’t fixable.