The next phase of artificial intelligence is taking shape inside the companies building it. AI agents are being trained to handle hours-long assignments, while researchers are using AI to write code, investigate technical problems and help develop the next generation of AI systems. AGI generally refers to an AI system capable of learning, reasoning and applying knowledge across a broad range of tasks, rather than being limited to a narrow set of functions.

There is no universally accepted definition or test for AGI, so there is still no agreed threshold for declaring that a system has reached it. But many jobs gradually moving into automation, pushing thousands into uncertainty might help us understand the whole story.

OpenAI says more than 70% of sampled Codex users in May 2026 asked the coding agent to handle tasks estimated to take more than an hour. About a quarter made at least one request estimated at more than eight hours. Anthropic reported in August that Claude was leading 26% of the AI research and development work covered by its internal measurement system, compared with less than 1% in February. Both figures come from the companies themselves and are not measures of the wider economy.

For most people, AI is still something they consult. Ask a question, get an answer. Give it a document, get a summary. Ask for code, get code. That model is changing. The newer systems can take an assignment, break it into steps, use software and other tools, check their progress and continue working with less human intervention.

What happens when AI can handle much more of the work itself? That question sits at the heart of the debate over artificial general intelligence, or AGI.

From Prompts to Assignments

A programmer can ask AI to write a function. An agent can be given a larger job: inspect an existing project, build a feature, run tests, find problems and make corrections. The AI is handling a sequence of tasks rather than producing one answer. That distinction could eventually change how many kinds of knowledge work are organised.

AI is Helping Build AI

Frontier AI companies are already making the change clear. OpenAI says it has developed an “automated research intern” capable of performing defined research tasks under human direction. The company says it is working towards an automated AI researcher by March 2028.

Anthropic’s August figures point in the same direction. The company says Claude is increasingly being used in its own AI research and development, although it remains dependent on human researchers and is not fully autonomous in the measured work. AI is helping researchers build better AI.

So, When does AGI Arrive?

Researchers disagree about the capabilities AGI should demonstrate, and no accepted test exists. Google DeepMind CEO Demis Hassabis said in May that he expected AGI could arrive within roughly four years, possibly sooner. If an AI can research a subject, analyse data, write software, use several digital tools and complete a complex assignment, how much of that work still needs to be done by a person? Work will change before we have an answer.

A researcher could delegate a literature review and data analysis. A programmer could hand over an entire software feature. Some tasks may disappear from jobs. Others may become faster. New work will emerge around supervising, testing and governing AI systems.

India is Preparing for the Shift

India is investing heavily in the infrastructure needed for the next phase of AI. The IndiaAI Mission has an approved outlay of ₹10,371.92 crore over five years. The government said in August 2026 that more than 45,000 GPUs had been onboarded through its shared computing programme and 237 projects had accessed subsidised capacity. Twenty indigenous foundation-model proposals had also been selected from 506 applications.

AI systems need to work across Indian languages and very different economic and institutional settings. OpenAI announced an India initiative with Tata Group in February covering areas including AI infrastructure and local capability. Google DeepMind has partnered with Indian institutions on applications in science, education, agriculture and energy.

It becomes harder when an AI system spends hours researching, writing code, analysing information and making decisions before presenting a result. The human role will increasingly involve setting the objective, judging evidence and deciding when a system should be trusted. AGI remains undefined, and nobody can put a reliable date on its arrival. But the shift towards more autonomous AI is already visible.

India’s AI ambitions are now reflected in global rankings, investment flows and Bengaluru’s emergence as a lead ing innovation hub. Yet the numbers reveal only part of the story. Much of the hardware is imported, demand for skilled workers continues to outpace supply, and the infrastructure pow ering AI faces growing pressure. The country’s next challenge is no longer adopting AI, but building the industrial and social foundations that can sus tain its growth.

The Hardware Production Gap

The IndiaAI Mission has deployed over 38,000 GPUs and TPUs across domestic data centres. Registered startups and researchers can access this compute at ₹115–150 per GPU hour, roughly 42% cheaper than com mercial cloud rates. The governmentplans to scale this to 100,000 GPUs by December 2026. Not one of these chips is made in India. Every GPU in that stack is sourced from the same export controls that cut China off from this hardware exist as legal authority that could, under different political conditions, apply to India too.

A US India trade framework announced in February this year includes language protecting India’s chip access, but it depends on ongoing political align ment. Budget 2026–27 allocated ₹8,000 crore to the semiconductor and display manufacturing ecosystem programme, the largest single-year outlay since the mission launched, with a separate ₹1,000 crore for India Semiconductor Mission 2.0. The Tata PSMC fabrication plant at Dholera is targeting trial production by late 2026. But these plants are not building AI grade chips.

India’s Sovereign AI

At the IndiaAI Impact Summit 2026, three models were introduced: Sarvam AI, BharatGen, Gnani.ai. BharatGen has assembled over 15,000 hours of annotated voice data across 22 Indian languages. Sarvam’s Vision model, a 3-billion-parameter doc ument intelligence system, scored 84.3% on a standard OCR benchmark, outperforming Google Gemini 3 Pro (80.2%) and OpenAI’s GPT 5.2 (69.8%). Bhashini-v2, launched in early 2026, offers AI-powered translation across all 22 scheduled Indian languages and serves 140 million users on the MyGov platform. India is constructing the language of sovereignty. Indigenous models, na tional compute, and mission branding. But the engine underneath runs on hardware it cannot make and may not always be allowed to import.

The Jobless Economic Growth

India’s IT sector has long been the engine of middle-class stability, con tributing about 7.3% to GDP, employing over 5.8 million people directly, and creating the white-collar jobs that drove spending across housing, edu cation, and retail. That engine is under pressure from the same technology India is racing to lead. Particularly for mid-level coding, BPO, and testing roles that form the bulk of jobs in India. TCS announced 12,000 layoffs in 2025; Infosys and Wipro followed. Across the sector, jobs have been cut in what companies call strategic realignment.

NITI Aayog’s October 2025 report projected that in a worst-case sce nario headcount could fall to 6 million by 2031. In April 2026, global equity re search firm Bernstein wrote an open letter to Prime Minister Modi warning that India’s 10 to 15 million IT services, GCC, and BPO workforce faces direct exposure to AI-driven automation. Sonal Varma, chief economist for India and Asia ex-Japan at Nomura, said: “Entry-level routine jobs are being displaced, and mid-level jobs are transforming. India needs to create about 8 million jobs annually.’’

Skill Set for the Emerging Sector

AI DevOps engineers, data centre operators, ethical AI auditors are the major emerging roles. But they require skill sets entirely different from the mass-hiring model that built the IT sector. An IIM-Ahmedabad study found that 68% of white-collar work ers fear automation within five years; 55% have adopted AI tools, but only 48% have received any training. India produces over 1.5 million engineering graduates annually but ranks 18th globally in skills alignment and 73rd in human capital in terms of AI. Graduate volume is not the same as workforce depth.

In that way adoption without re skilling becomes exposure to disrup tion dressed up as progress. NASSCOM projects that the broader AI push could generate 750,000 jobs and add $500 billion in economic value by 2030. Whether they reach the workers being displaced depends on whether India can close the gap between how many people it trains and how well it trains them. It is a gap that every major index in 2026 has documented and none has resolved.

The Climate Bill

There is a cost to India’s AI ambitions that doesn’t appear in invest ment announcements is the physical climate risk accumulating around the infrastructure meant to run it all. A report by climate risk consul tancy assessed 2,595 planned data centres worldwide, examining risks of direct physical damage from climate hazards, operational disruption from extreme heat.

For India, the country ranks 11th globally in physical climate risk to planned data centre infrastruc ture. The more concerning finding is where that risk concentrates. Tamil Nadu, Telangana, and Karnataka, three of India’s major data centre corridors are among the top 30 regions globally for projected operational disruption from extreme heat. South Asia as a whole has one of the highest pro portions of high-risk planned facil ities globally, with risk projected to increase sharply toward the end of the century.

Data centres require large-scale cooling to keep servers running. Rising ambient temperatures increase cooling costs, strain electricity grids, and raise the probability of outag es. Countries including India, Brazil, Mexico, Indonesia, and Spain already record some of the highest projected operational disruption risks from heat globally, with more than 75% of anal ysed facilities classified as high risk. Productivity losses become ten times higher with indirect risks like power outages, water shortages, infrastructure failures.

India’s data centre ecosystem is concentrated in heat-exposed regions with al ready-strained urban infrastructure. Future vulnerability can be reduced by adequate planning during site selection, engineering standards, and resilience investment. Microsoft has committed $17.5 billion to Indian data centre expansion. Google is building a hub in Andhra Pradesh with a $15 billion commitment. Amazon has pledged $48 billion through 2030. At this stage, planning choices that are made now determine the risk profile for the next three decades. India has no mandatory national standard governing data centre siting or construction for climate resilience. This gap is no longer a future problem.

Where the Momentum Meets Reality

The 190% growth in Bengaluru’s ecosystem, the indigenous models outperforming global benchmarks, the public compute infrastructure being built at scale are all substantiated achievements. The QS Index gives In dia a perfect score of 100 on economic capacity and ranks it fifth globally in the “Future of Work” category. “The IndiaAI Mission represents substan tive, measurable progress, not merely a policy optics exercise.”

But the gaps are structural. The top ten Indian cities account for nearly half of all AI users while representing less than 10% of the population. The choices being made about who controls the technology, who gets the jobs, who bears the risks will deter mine whether this moment becomes something real or just another set of promises that get quietly shelved. What happens in policy rooms and planning offices will have to answer it.

AI agents are beginning to do more than answer questions. They can browse the internet, use software, interact with external systems and act on a user’s behalf. That greater autonomy is also producing a new AI safety category of failures: situations in which an AI system does something its user did not intend or cannot easily control.

A new research registry published this month documents a growing body of reported incidents involving AI agents and distinguishes real-world failures from controlled safety demonstrations. Its authors found that cases involving actual harm are concentrated among incidents occurring in the wild and those linked to safety failures.

AI Safety and People Building the Systems.

Former Google DeepMind research engineer Bilal Chughtai warned this week that advanced AI could pose an extreme threat to humanity. His resignation in July was followed by public warnings from Jacob Coxon, a former researcher at Anthropic and OpenAI. At Anthropic, researcher Evan Hubinger has estimated that there is a greater than 10% chance of a catastrophic AI outcome within the next decade.

Anthropic CEO Dario Amodei has called for a slowdown in frontier AI development, while OpenAI CEO Sam Altman has also argued for greater coordination around increasingly capable systems.

Against this backdrop, a team of MIT researchers is tackling a narrower but increasingly relevant problem: how do you give generative AI enough freedom to find good solutions without allowing it to violate rules that cannot be broken?

When Almost Right is Not Enough

For a generative model, a plausible answer is often considered a success. That standard does not work when the output controls a robot, a physical process or another system with strict operating limits. A robot planning a route through a factory, for instance, cannot simply find a path that is unlikely to collide with a worker. It has to avoid the collision. These non-negotiable requirements are known as hard constraints.

One way of enforcing them is to constrain the model repeatedly as it generates an answer. But that can narrow its search too early, preventing it from finding a better solution. MIT researchers Zeyang Li, Kaveh Alim and Navid Azizan have developed HardFlow, a method that takes a different approach.

Instead of forcing every intermediate step to satisfy the constraints, HardFlow allows the model more freedom while it searches and steers the generation process towards a final output that meets the required conditions. The researchers frame this as a trajectory-optimisation problem, drawing on optimal control theory.

They also break the computational problem into smaller steps so the method can be used at deployment time without retraining the underlying generative model. In experiments involving robotic manipulation, maze navigation, physical-process control and text-guided image editing, HardFlow satisfied the specified constraints while producing higher-quality solutions than competing methods.

The significance is narrower than the wider AI safety debate, but that may be precisely what makes the work useful. HardFlow does not attempt to solve every problem associated with increasingly autonomous AI. It addresses a practical question: how can a model retain the freedom to search for a good answer while ensuring that the final answer stays within boundaries that cannot be negotiated?

As AI systems gain greater autonomy, that distinction may become increasingly important. The challenge is not simply to make AI more capable, but to ensure that capability operates within limits that remain under human control.