ChatGPT can now do your work without you. Agent mode is here, faster than expected, and with it, a very different future to the one a lot of people expected. Are consumers ready? Are you ready?

OpenAI has launched ChatGPT Agent Mode, a unified system that lets GPT control a virtual computer, browse the web, and operate a command line to complete tasks for you.

ChatGPT has moved beyond chat. Agent Mode marks a fundamental shift from reactive assistant to autonomous operator. Internal OpenAI tools once siloed (Operator, Deep Research, and ChatGPT) are now fused into a single system that handles real work: generating reports, editing spreadsheets, summarising complex files, even navigating the web. Confirmation appears only when required, such as sending emails, making purchases, or accessing accounts. “Watch Mode” activates during sensitive sessions to ensure transparency without blocking execution. Users provide the goal. ChatGPT manages the strategy.

Agent Mode runs on three core systems: 1) The Browser provides a lean, text-based search engine with full-page navigation, optimised for results rather than aesthetics, 2) the Computer functions as a virtual desktop with standard productivity tools including LibreOffice, file explorers, and PDF readers, and 3) Container offers a secure, sandboxed terminal designed for scripting, debugging, and advanced developer operations.

The result is a system capable of managing multi-step workflows, reasoning through complex instructions, and executing without micromanagement. Not just a smarter chat interface, more a foundational layer for autonomous work where delegation is the interface. As a result, Operator, the standalone product, is being deprecated. Rollout live for 100% of Pro users and Plus and Team users in the US will start getting access from Monday. Users in the European Economic Area and Switzerland have to wait. Usage limits are tiered with Pro users getting 400 messages per month (Plus gets 40, Team get 30). Only user-initiated actions that push the agent forward count towards these limits; clarifying questions and system confirmations do not. If that sounds annoying, weird and confusing it’s because it is. Need something simple? Here’s an example of it doing your shopping for you from Alex Banks:

Safety is understandably a huge area for tools like this and OpenAI made special mention of what the company has done to ensure safety from prompt injection defences, file and domain-based access controls, memory and input sanitisation, and a new system card that explains what Agent can and can’t do. Understandably, OpenAI have classified this as High Capability in biology and chemistry under its Preparedness Framework.

Also this week: OpenAI tested shopping checkout in ChatGPT, and hit gold medal–level performance on the 2025 International Math Olympiad, not with a formal maths engine, but with a general-purpose reasoning model. Both huge news in their respective industries, Sam Altman decided to add more fuel to the fire saying that GPT-5 was ‘coming soon’ (lots of speculation before the end of July) along with more details that GPT-5 will be built on ‘experimental research that will shape future models’, not an Olympiad-level version, that won’t ship for “many months.”

Be under no illusion, OpenAI just pushed agentic computing into the mainstream, quietly and quickly, without waiting to see how people might respond making it look cute and simple in the process.

For enterprise, this is not a feature release but a structural shift. ChatGPT Agent now has the capability to bypass entire workflows with a single instruction, cutting through UX layers, onboarding processes, procurement steps, and internal teams. A powerful system now sits between users and infrastructure, changing how control flows within organisations. Are they ready for it? Most are not. Power is beginning to shift away from traditional roles like managers or compliance leads, and towards those building the connectors, tuning the models, and shaping agent behaviour. Influence is being redistributed, but it remains unclear who will emerge with the most.

Early testers describe Agent Mode as simple to operate but difficult to steer, offering no insight into its process and no way to influence how tasks unfold. The lack of control may limit its appeal for advanced users and poses challenges for IT and security teams that require visibility. So far, usage has remained light; planning travel, generating decks, assisting with admin, but the direction is clear. Very few anticipated the technology arriving this soon, and even fewer have a plan for what comes next. The central question is not whether agents will take over, but what organisations are doing before they do. The agent age is here, and the realignment of influence has already begun.

Got this forwarded to you? Want to get it regularly? Smart cookie.
Choose monthly (£15) or annual (£99). Cancel anytime; no monsters here.

> Google more your thing? There’s a WD_DTW? for them too. Go here.

> Amazon more your thing? There’s one for them too. Here it is.