The era of the passive chatbot is ending. In its place, a new class of autonomous software — AI agents capable of navigating the web, manipulating interfaces, and executing multi-step workflows — is quietly moving from research labs into the browser tabs of early adopters. Meta’s Muse and OpenAI’s Dots are leading this shift, promising to handle the digital grunt work that consumes hours of modern life: tailoring CVs for specific roles, hunting for elusive restaurant reservations, and managing the administrative overhead of simply existing online. For the first time, the promise of a genuine digital assistant feels less like science fiction and more like a productivity hack.
The Shift from Generation to Action
For two years, the generative AI narrative has centred on creation: writing code, drafting emails, summarising PDFs. But the industry’s frontier has moved. The new metric of value isn’t token output; it’s task completion.
Muse and Dots represent a architectural pivot. They are not merely large language models wrapped in a chat interface. They are systems designed for agency — the ability to perceive a digital environment, plan a sequence of actions, and execute them with minimal human supervision. Think less “write me a cover letter” and more “find 20 product marketing roles in London, tailor my CV for each, draft bespoke cover letters, and submit the applications before lunch.”
This requires a fundamentally different stack. Under the hood, these agents combine multimodal perception — reading the DOM of a webpage, interpreting visual layouts — with reinforcement learning techniques honed on tool use. They browse, click, scroll, type, and recover from errors. When a “Apply Now” button shifts position after a site redesign, the agent adapts. When a login wall appears, it pauses for credentials. It is the difference between a knowledgeable intern and a competent executive assistant.
Five Workflows Where the ROI Is Immediate
Early users are converging on a handful of high-friction, high-volume tasks where the time savings are undeniable.
Job hunting at scale. The most cited use case. Agents can ingest a master CV, scrape LinkedIn, Greenhouse, and Lever boards for matching roles, rewrite the resume to mirror the job description’s keywords, generate a tailored cover letter, and hit submit. One beta tester reported 47 applications completed in a single evening — a volume that would have taken a human weeks.
The reservation hunt. Securing a table at a Michelin-starred restaurant or a trendy Saturday brunch spot often requires refreshing a booking widget at the exact second slots drop. Agents monitor availability in real-time, solving CAPTCHAs and navigating multi-step confirmation flows faster than any human reflex allows.
Comparative shopping across fragmented markets. Need a specific vintage watch, a sold-out GPU, or the cheapest business-class fare with a lie-flat seat? Agents can crawl dozens of retailers, aggregators, and forums simultaneously, normalising data into a single comparison table — including shipping, tax, and return policies.
Administrative triage. Renewing driving licences, updating address details across twelve vendors, cancelling unused subscriptions buried in email receipts. These are death-by-a-thousand-cuts tasks. Agents excel at the repetitive navigation of brittle government portals and legacy corporate account centres.
Research synthesis with citations. Not just summarising, but verifying. Agents can be instructed to “find every peer-reviewed study from the last three years on GLP-1 agonists and cardiovascular outcomes, extract the hazard ratios, and build a spreadsheet with DOIs.” They browse PubMed, download PDFs, parse tables, and structure the output. The human role shifts from gatherer to analyst.
The Trust Barrier and the “Human in the Loop”
Autonomy breeds anxiety. Handing over credentials to a non-deterministic system is a hard sell for security teams and privacy-conscious individuals. Both Meta and OpenAI have architected their agents around a “human-in-the-loop” paradigm for sensitive actions.
Payments, logins, and final submissions require explicit confirmation. The agent prepares the cart, fills the forms, reviews the terms — then pauses. “I’m ready to submit this application. Shall I proceed?” This checkpoint model builds trust while preserving the bulk of the time savings. It also creates an audit trail: every click, every keystroke, every decision node is logged and replayable.
There is also the question of identity. When an agent applies for a job on your behalf, is it you? Recruiters are already detecting AI-generated applications at scale. The arms race between agent-generated submissions and AI-detection filters has begun. Some platforms now require video introductions or live coding challenges precisely to verify human agency. The irony is palpable: we build agents to navigate systems designed by humans, only to force humans to prove they aren’t agents.
The Platform Play: Walled Gardens vs. Open Web
Meta’s strategy leans into its ecosystem. Muse integrates deeply with WhatsApp Business, Instagram DMs, and the broader Family of Apps. Its superpower is social context — it knows your network, your messages, your calendar. OpenAI’s Dots, by contrast, is positioning as a neutral layer atop the open web, accessible via API and the ChatGPT interface, agnostic to the underlying service.
This divergence matters. A Muse agent might seamlessly coordinate a group dinner by polling friends on WhatsApp, checking calendars, and booking the table. A Dots agent might be better suited for the solitary, cross-platform research tasks that span competitors’ territories. Developers are already building wrappers that orchestrate both: Muse for the social-commercial layer, Dots for the open-web heavy lifting.
Why it Matters
We are witnessing the birth of a new labour category: digital labour that is cheaper, faster, and infinitely scalable than human virtual assistants, yet more capable than rigid RPA scripts. The immediate impact is personal productivity, but the structural implication is a rewriting of the internet’s incentive structures. When agents mediate our interaction with websites — scraping, filtering, transacting — the user interface becomes an API. Ad-supported models, dark patterns, and attention-engineered designs lose their leverage. The web was built for human eyes; the agentic web is built for machine readability. The companies that adapt their digital surfaces for agent consumption — clean markup, structured data, programmatic checkout — will capture the next wave of commerce. Those that don’t will find their customers’ agents simply bypassing them.
