AI agents start acting like coworkers, not tools

Autonomous AI agents are no longer lab toys or browser tabs—they’re booking your Uber, planning your commute, and quietly running campaigns from a spare Mac mini while you sleep. That shift, from “assistant you prompt” to “coworker that acts,” is beginning to redraw who actually executes digital work and how much of it humans still touch.

The first wave is arriving on consumer phones. At Samsung’s latest Galaxy launch, Google and Samsung showed how Gemini can now open third‑party apps like Uber, DoorDash, and Grubhub in a secure virtual window, navigate screens, fill in forms and confirm payments, all from a single voice command such as “Get me an Uber to the airport” or “Order this on Grubhub for home delivery,” while you keep using the phone for other tasks, as detailed by Wired. Parallel reporting notes that the feature is rolling out first to Pixel 10 and Galaxy S26 devices, with tasks executed in an on‑device sandbox and surfaced via notifications you can pause or cancel, a set of guardrails Google outlined in an interview covered by Yahoo/Android Update. This is not just voice control; it is end‑to‑end task execution across apps.

Maps that answer—and then act

Google is also folding these agentic abilities into one of its most widely used interfaces: Maps. A major redesign announced this week adds an “Ask Maps” conversational layer that lets users describe intents like “plan me a road trip with dog‑friendly stops” and receive tailored routes, recommendations, and itineraries, grounded in data from more than 300 million places and 500 million contributors, according to an overview from the Associated Press. For now, Ask Maps focuses on planning and Q&A, but Google is already exposing the same geospatial data and tools through its Gemini API so developers can wire agents directly into location‑based workflows, as explained in a recent Google developer blog.

The combination—agents that can both reason about the physical world and click through apps that affect it—turns Maps from a passive directory into something closer to a dispatcher. Third‑party platforms are moving in the same direction: workflow tools like Arahi and Tars pitch “AI agents with Google Maps” that can trigger automations when location data changes, embed live maps in chat responses, or chain geospatial events into broader business workflows, as described in their integration guides on Arahi and Tars. Instead of a human scheduling a site visit or updating a field technician’s route, an agent can increasingly do it on its own.

From no‑code builders to 24/7 digital staff

On the enterprise side, startups are racing to put these agents to work at scale—and to hand the keys to non‑technical employees. Gumloop, a no‑code AI automation framework, markets itself as a way for “any team member” to turn manual processes into automated workflows without writing a line of code, a promise the company highlights on its own product page. Analysts tracking the company estimate it has raised more than $20 million to date, including a sizable Series A, to push this vision of “every employee as an agent builder,” according to a teardown from NextLeap. The bet is that line‑of‑business staff—not central IT—will own and iterate on their automated coworkers.

Perplexity, meanwhile, is pushing agents out of the browser and into the local network. At its first developer event this week, the company announced “Personal Computer,” a product that runs as an always‑on AI controller on a Mac mini, linking local apps and files to Perplexity’s cloud models so an agent can operate 24/7 on behalf of a team. Coverage from The Deep View describes it as turning Perplexity “from a browser‑based agent into a 24/7 proactive AI running on a Mac mini,” aimed at use cases like continuously tuning ad campaigns or monitoring analytics dashboards, positioned as a rival to open‑source setups like OpenClaw The Deep View and summarized with pricing details by industry outlet The Outpost. In marketing pitches and user testimonials, teams describe these agents as replacing stacks of SaaS tools rather than just adding another one.

New productivity—plus new attack surface

In theory, these agents collapse busywork: no more tab‑hopping to schedule deliveries, triage inboxes or copy‑paste analytics into reports. Early Perplexity products, like its AI “Email Assistant” for Outlook and Gmail, already show how an agent can categorize mail, draft responses in a user’s tone, and propose calendar slots for meetings, albeit behind a steep $200‑per‑month paywall that Windows Central notes effectively limits adoption to high‑value professional users.

But putting an autonomous system in charge of app clicks multiplies risk. Each new agent that can read group chats, authorize payments or browse internal dashboards becomes a high‑value target for prompt injection, social engineering and traditional malware. Security researchers have already flagged how browser‑integrated agents like Perplexity’s Comet blur the line between user intent and automated actions across tabs, a dynamic TechRadar highlighted in its coverage of Comet’s cross‑tab multitasking update, which emphasized the need for explicit permission prompts before the AI takes action in the browser TechRadar.

The labor implications are just as immediate. If a marketing “stack” or customer‑support queue can be turned into a Mac‑based agent that runs around the clock for a few hundred dollars a month, managers will inevitably compare that cost to salaries, benefits and training. At the same time, organizations will need new roles—people who design these workflows, monitor agent behavior, and audit decisions for bias, error and abuse. In other words, your new coworker may be an AI agent that never sleeps, but your new job might be making sure it doesn’t quietly rewrite how your organization works without anyone noticing.