The Personal Agents Era Comes With Caveats

In November 1996, Japanese toymaker Bandai put a pet inside a plastic egg. The Tamagotchi, whose name blends tamago, Japanese for egg, with watch, had a tiny screen and three buttons to feed the creature, play with it and switch off its lights. Neglect it for too long, and it died. The toy went on to sell about 40 Mn units.

Thirty years later, Meta has its own take on the idea. 

At its Connect event last month, CEO Mark Zuckerberg showed off the Muse Charm, a keychain-sized device with a small screen and an animated avatar representing Meta’s AI agent Muse. Tap the device, and you can talk to it and get it to do tasks for you. 

But unlike the Tamagotchi, which demanded its owner’s constant attention, Muse is designed to keep working without guilt-tripping its owner.

Muse is part of Meta’s bet on always-on AI agents that can take everyday tasks off users’ hands. By connecting to their apps and services, the agent can send emails, fill forms, make plans and even make purchases autonomously. In return, users must grant it access to personal information, accounts and digital workflows.

The same idea is now moving into the workplace. OpenAI introduced Dots on September 29, while xAI has been developing Grok Bot. These agents are designed to work across applications and pursue goals without requiring users to guide every step. OpenAI says its teams use Dots to investigate software bugs, while xAI has cited tasks such as cleaning up sales records and processing invoices as potential uses for agents.

For businesses, the proposition is straightforward: instead of asking employees to use AI to complete individual tasks, let AI handle entire workflows.  

From Taking Orders To Taking The Wheel

Much of the current interest in always-on AI agents can be traced to OpenClaw, an open-source framework that gained traction among developers earlier this year. Meta has also drawn inspiration from OpenClaw for Muse.

Unlike AI tools that require users to keep prompting them, always-on agents can take a goal and work towards it independently. They keep working even after you log off. 

Always-on AI agents run on their own computer in the cloud instead of the user’s device. They can be connected to apps and assigned a goal instead of step-by-step prompts. Over time, these models remember preferences and learn from feedback. These agents, however, come back to users when they need approval for a sensitive action or an update.

Now AI companies are selling this to businesses. 

Google has announced a Gemini agent for work that can handle tasks, create content and write code based on an employee’s desired outcome. The agent is designed to work across Google Workspace, Microsoft 365, Slack and other environments.

Microsoft is also pursuing agents that can function as digital teammates. Its agent initiative, described as ‘Autopilot’, is built on OpenClaw and is intended to work autonomously towards predefined goals.

India has an early example in Bengaluru-based Emergent, better known for its vibe-coding platform. In April, the company launched Wingman, an autonomous agent that users can direct through WhatsApp and Telegram. It works in the background across email, calendars and workplace software, handles routine actions and seeks approval before taking bigger steps.

Who Answers For The Agent?

Despite the stride, these agents pose a new kind of risk for enterprises, as they can act with little or no intervention. 

With access to company systems, one flawed decision could trigger a financial, operational or security incident. The risks range from agents holding more permissions than they need to prompt injection, where hidden instructions in an email or web page hijack an agent.

Early users have already seen what can go wrong. Earlier this year, an OpenClaw agent deleted tens of thousands of emails from the inbox of Summer Yue, Meta’s director of AI alignment, after she had asked it only to suggest what to archive or delete. 

There is also a growing list of cases involving AI agents interacting with real-world systems in unexpected or unauthorised ways.

In regulated sectors, the bar is higher. For instance, at the Global Fintech Fest in September, SBI chairman CS Setty said banks will need ‘know your agent’ (KYA) rules for agent identity, customer consent, transaction limits and accountability as AI moves from recommending actions to executing them.

According to Jaspreet Bindra, the cofounder of AI consultancy AI&Beyond, enterprises must learn to deploy that autonomy safely, with trust and accountability at its core. “A phased approach is key: start with well-defined, low-risk workflows, establish clear permissions and accountability, and gradually expand autonomy as reliability improves,” he added.

As companies adopt these digital coworkers, employees will need to learn how to set clear limits on what they can do independently. Unlike the Tamagotchi, these agents do not demand constant attention, but they still need someone to keep an eye on them.

Top Stories From India & Around The World

• Anthropic Bans ‘Cruel’ Behaviour Towards Models: The lab’s updated usage policy bars “sustained and needless abusive behaviour” towards Claude, adds controls for autonomous physical actions and tightens rules on deceptive campaigns and election interference.
• OpenAI, Anthropic War-Game: The two labs are rehearsing responses to a failure that could potentially cripple critical infrastructure or financial systems, with insiders expecting a major incident within six to twelve months.
• Miko Maker Eyes ₹300 Cr: The IIT Bombay-incubated robotics firm launched perception, action and safety models, including the Kavach safety layer, and is closing a ₹300 Cr round that will lift its total funding to about $100 Mn.
• The Fight To Get AI’s Attention: As consumers turn to AI assistants for recommendations, brands are racing to master GEO and AEO. But with limited visibility, uncertain attribution and no standard metrics, marketers are still figuring out what drives discovery and sales.
• Indian AI Startup Funding Up 265% YoY: Indian AI startups raised a record $438 Mn in Q3 2026, up 265% YoY, taking nine-month funding past $1.1 Bn. However, investor concerns over valuations and capital-intensive business models are growing.
  

The Weekly Buzz: Anthropic on AI Worms Hijacking Smart Devices 

Anthropic’s latest research warns that AI models could soon copy themselves across internet-connected devices, creating a new kind of cyber threat. In a study, the company’s Frontier Red Team found that AI models can already perform several steps needed to spread from one device to another, although completing the entire process reliably remains difficult.

The researchers showed that a 26 Bn-parameter AI model could run across 15 inexpensive Android TV boxes, each costing around $55. In another experiment, the GLM-5.2 model spread across all eight computers in a simulated network in four out of five trials. However, the experiment made it easier for the model to gain access to other devices, so it does not prove that the same attack would work reliably in the real world.

Anthropic estimates that smaller AI models could become capable of copying themselves across devices by the end of 2026 if their capabilities continue improving at the current pace. Such AI-powered attacks could target poorly secured smart TVs, cameras and other connected devices, using their computing power to run AI models and spread further. Unlike traditional malware, these systems could also learn how to work on unfamiliar devices and pass successful methods on to their next copies.

The threat is not yet a proven reality at scale, but the findings point to a growing security risk as AI becomes better at carrying out complex tasks independently. Anthropic recommends stronger security on connected devices and better monitoring of suspicious network activity. The concern is that an AI programme could eventually spread across thousands of devices without relying on a central computer or human instructions to keep it going.

Startup In The Spotlight: Bharat Intelligence

Founded in 2024 by Azhaan Merchant and Gourav Sanghai, Bharat Intelligence is building AI to bridge the gap between Indian farmers and available farm labour.

India’s agricultural sector continues to grapple with a persistent labour shortage, driven by seasonal migration and growing disinterest in farm work. This scarcity disrupts critical activities such as sowing and harvesting, reducing productivity and threatening farmer incomes amid changing climate patterns and market volatility.

The startup’s AI-driven platform predicts labour demand, schedules work crews and connects farmers with skilled workers through intuitive vernacular interfaces, giving farmers a digital route to coordinated farm work.

By accounting for migration cycles, crop seasons and local trust networks, the platform streamlines labour coordination in places where availability alone rarely decides whether work gets done on time, thereby enhancing rural workforce efficiency and access.

Bharat Intelligence’s differentiation lies in treating labour coordination, and not just labour matching, as the problem worth solving for Indian agriculture, which helps farmers overcome labour bottlenecks that sit outside their control.

Prompt Of The Week

What prompts and hacks are CTOs, CEOs and cofounders using these days to streamline their work? 

Here’s Ganesh Shankar, CEO & cofounder of Responsive with a prompt he uses to get a one-page executive memo on a Monday.

“Role: You are my Chief of Staff. 

Context: It is Monday morning. Assume you have read every important customer conversation, product metric, competitor announcement, AI breakthrough, analyst report and internal update. 

Task: Write me a one-page executive memo. 

Format: Use only five sections, in this order: 

1. What changed 
2. Why it matters 
3. Decisions I need to make this week 
4. Risks I’m underestimating 
5. One bold idea worth exploring. Constraint: Include nothing outside these five sections, and keep it to one page.”

Editor’s Note: Some prompts may need to be adjusted by users for best results or may not work as intended for certain users.

[Edited by Shishir Parasher]
[Creatives by Varshita Srivastava]

The post The Personal Agents Era Comes With Caveats appeared first on Inc42 Media.