The era of the passive chatbot is effectively over. This week, both Meta and OpenAI pushed significant updates that move their artificial intelligence models from conversational partners to autonomous operators, capable of navigating the open web to complete complex, multi-step chores. Meta introduced Muse, a research agent designed to synthesise vast amounts of online data, while OpenAI began rolling out “dots”—a codename for agentic capabilities inside ChatGPT that can execute tasks like filling out recruitment portals or wrestling with reservation systems. For the first time, mainstream users can delegate genuine digital labour to an algorithm, provided they are willing to watch it work.
The Shift from Generation to Action
For eighteen months, the generative AI narrative has centred on creation: writing copy, debugging code, summarising PDFs. The new paradigm is agency. Muse and dots represent distinct architectural bets on how software should interact with the internet on a human’s behalf. Meta’s approach leans heavily on its massive search index and the Llama 3 foundation, positioning Muse as a deep-research analyst that reads hundreds of sources, cross-references claims, and delivers a cited report in minutes. OpenAI, conversely, is betting on the browser environment itself. Its agents drive a virtual cursor, clicking buttons, typing into form fields, and scrolling through dropdown menus exactly as a user would—only faster and without the fatigue.
The distinction is subtle but critical. Muse retrieves and reasons; dots navigates and executes. Both require a human in the loop for final verification, but the cognitive load has shifted dramatically. Instead of prompting, refining, and prompting again, the user defines an objective—“find me a senior product role in fintech paying over £100k” or “book a table for four at The Wolseley next Friday at 7:30 pm”—and the agent manages the workflow.
Putting the Agents to the Test
Early access users and developers have spent the last seventy-two hours stress-testing these capabilities across four high-friction domains. The results suggest the technology is ready for prime time in specific, structured scenarios, though it still stumbles on ambiguity.
Job applications have emerged as the killer use case. Dots can ingest a CV, scan LinkedIn and corporate career pages for matches, tailor cover letters to specific job descriptions, and submit applications through Greenhouse, Lever, and Workday portals. One tester reported submitting forty tailored applications in the time it usually takes to complete two. Muse complements this by researching the target company—earnings calls, Glassdoor reviews, recent press—to brief the candidate before the interview stage.
Restaurant reservations are the second high-value target. Securing a table at sought-after venues often requires refreshing a booking widget at the exact second slots drop, or navigating phone systems that refuse to connect to a human. Agents handle the timing and the navigation. They can monitor Resy, OpenTable, and proprietary systems, alerting the user only when a confirmation is secured. For popular London spots like Brat or Kiln, this capability alone justifies the subscription fee for many early adopters.
Travel itinerary assembly is the third domain. Agents can now compare flight matrices across carriers, cross-reference hotel cancellation policies, check visa requirements for specific passports, and build a day-by-day schedule in a shared document. They struggle, however, with subjective preferences—“a vibey neighbourhood with good coffee”—unless the user provides explicit parameters.
Competitive intelligence gathering rounds out the quartet. Sales teams and founders are using Muse to monitor competitor pricing changes, product launches, and hiring signals across hundreds of web pages daily. The agent flags anomalies—a sudden spike in hiring for AI roles at a rival, a price drop on a key SKU—and delivers a daily briefing before 8 am. This transforms market monitoring from a weekly manual sweep into a continuous, automated feed.
The Guardrails and the Gaps
Neither company is claiming full autonomy. Both systems operate within “sandboxed” browser environments, and sensitive actions—payment confirmation, sending a final email, accepting terms of service—require explicit user approval. OpenAI has implemented a “takeover” mode where the user can seize control of the virtual mouse at any moment. Meta has restricted Muse from interacting with private, logged-in accounts, limiting it to public data for now.
Hallucination remains a risk, though the nature of the error has changed. A chatbot invents a fact; an agent clicks the wrong button on a live website. In testing, dots occasionally selected the wrong date on a calendar widget or misread a salary field as monthly rather than annual. These are not catastrophic failures, but they demand supervision. The productivity gain comes from offloading the mechanical ninety percent of the task, not the judgement critical ten percent.
Privacy frameworks differ sharply. OpenAI states that browser sessions are ephemeral and not used for model training. Meta’s Muse operates under the company’s broader data policy, which has drawn scrutiny from European regulators. Enterprise customers are advised to run evaluations through their own data governance boards before deploying agents on proprietary internal systems.
Why it Matters
We have crossed the threshold from AI as a tool you use to AI as a colleague you manage. This is not merely a feature upgrade; it is a fundamental restructuring of digital labour. When the marginal cost of executing a web-based workflow drops near zero, the bottleneck moves from execution to intent definition. Professionals who master the art of prompt engineering—specifying objectives, constraints, and success criteria with precision—will outpace those who treat these agents as magic boxes. The winners in this next phase will not be those with the best models, but those with the clearest processes.
