Table of contents
Launched in mid-July 2025, ChatGPT Agent represents OpenAI's ambition to take a step towards autonomous assistants capable of navigating the web, executing scripts, and completing complex tasks end-to-end. Presented as an “on-demand workforce,” the tool generates both enthusiasm and frustration, according to feedback from early users.
Real Capabilities but Mixed Performance
On paper, ChatGPT Agent impresses with its ability to chain the steps of a mission: online information gathering, organization and synthesis, document generation (reports, presentations, spreadsheets), the tool combines the functionalities of Operator and DeepResearch, already available on the platform for several months. For structured and unambiguous tasks, several testers report significant time savings, with the agent managing to cut the time needed for certain research and deliverable preparations by half or more.
However, the limits quickly appear when tasks become more complex. Numerous reports on Reddit and LinkedIn highlight noticeable slowdowns, repetitive action loops, and a high failure rate on certain web interactions.
The Sandbox: Security or Constraint?
The operation of ChatGPT Agent relies on a sandbox: a virtual environment hosted by OpenAI, integrating a browser and a virtual desktop. This approach aims to limit security risks and maintain control over the agent's actions, but it significantly hinders its capabilities.
Many sites block access or detect the agent as a bot, preventing reservations, purchases, or complex interactions.
The agent regularly fails with dynamic forms, highly interactive sites, protected by CAPTCHA or services like Cloudflare... rendering it unusable for fulfilling most of the most enticing promises on paper.
This protective approach makes the agent usable for research and synthesis but unreliable for advanced or critical web actions.
An Attempt to Catch Up with Manus
In contrast to ChatGPT Agent, Manus, developed by Future AGI, offers a more daring approach. Using a multi-agent architecture (planning, execution, validation) and a more comprehensive sandbox environment (advanced browser, terminal, multimodal generation), Manus is perceived by some as more efficient on complex tasks and autonomous workflows.
However, Manus remains less accessible (restricted access, high credit costs, limited availability), which limits its adoption. User feedback reports frequent bugs and significant credit consumption in case of failure, generating frustration. Its autonomy implies less real-time control, which can lead to deviations when the agent goes in an unexpected direction.
Towards Agents Integrated into the Browser?
Another path could alleviate some of the limitations: relocating the agent into the user's browser, via an extension or a local module.
Such a model would offer:
- Faster execution by eliminating virtualization.
- Direct integration with local tools and data (CRM, messaging, documents).
- Better compatibility with modern sites, as it would behave like a real human browser.
But this choice would pose two major challenges:
- Security: an agent with access to cookies, accounts, and local data would become a prime target for abuse.
- Marketing positioning: a local agent would resemble more of a personal automation tool than an “outsourced workforce,” changing OpenAI's commercial promise.
A Product Still in Development
As it stands, ChatGPT Agent remains an innovative but immature tool: useful for speeding up certain well-defined tasks, limited for more ambitious missions. The vision of an “AI workforce” is more of a marketing promise than a technical reality.
The evolution of the AI agents market could involve a hybrid compromise: agents capable of working in a secure environment while delegating some interactions to the local browser, with increased control and safeguards.
Meanwhile, users should consider ChatGPT Agent as an intelligent assistant in testing phase, not as a true replacement for human collaborators.
What Future for the Web in the Face of Agents?
The internet has always enabled software and computer systems to communicate with each other. Most websites implement at least one or more APIs, interfaces for communication between software. Today, agents aim to transform the visible part of the web, usable by humans, into natural language software interfaces.
The question is not so much about technical capabilities as it is about whether publishers consent to making their content accessible to software. The theoretical promises of agents are based on a world where agents have access to everything. One of the first evident consequences would be the acceleration of the already ongoing collapse of the advertising model.
This could also raise questions about shifts, particularly in the case of marketplaces: how can we imagine that the good deals on sites intended for individuals, like Le bon coin or Vinted, are not completely pre-empted systematically by a few players?
What will happen to competition and diversity of offerings if we all use the same price comparator? Publishers are caught between circumventing their usage policy and becoming progressively invisible.