As more organisations turn to artificial intelligence (AI), specifically agentic AI, to improve their operational efficiencies, focus also has turned to the performance of the models that power AI workflows.
However, achieving scale alone may not be sufficient in improving AI models.
Instead, how models learn from interactions and update with context will be crucial, said Sara Hooker, CEO and co-founder of Adaption, an AI startup that focuses on building models that adapt to real-world conditions.
AI models must continue to learn as they operate so companies know what is, and is not, working, Hooker said, during a panel discussion at ATxSummit 2026 in Singapore.
It also narrows the gap between experimentation and real-world performance and impact, she said.
Where AI has fallen short is in compound decision-making, which is what AI agents are deployed to handle, she noted.
She added that automation scientists are focused on how to train models to update and learn, so they can have real-world impact.
Continuous learning is now paramount, she said, and key to frontier AI models.
AI learning from humans
Nvidia also is training its robots by letting the systems learn from humans, according to William Dally, the chipmaker’s chief scientist and senior vice president of research.
Beyond tapping visual natural language models, its robots are pretrained using video demonstrations of the same tasks being carried out by humans, Dally said during a fireside chat at the summit.
Actions are further finetuned through actions-based simulation, during which humans put on sensors-enabled gloves to train the robotics.
Nvidia believes simulation-based learning and training robots to autonomously perceive and act in dynamic conditions help improve the performance of the AI systems.
Preprogrammed robots are useful for specific and repetitive tasks, but they operate based on fixed instructions within set environments and are limited in their ability to adapt to unexpected changes.
AI-powered robots address such limitations through simulation-based learning and can acquire as well as refine new skills by leveraging learned policies, such as behaviours for navigation.
These can improve their ability to adapt and make decisions in various situations, hence, reducing the gap between simulations and real-world deployments.
The more training data goes into the robotics, the higher the accuracy in performing the tasks, Dally added.
AI agents also must have the right data, including company policies, to make the right decisions, said Sanjay Gupta, Google’s Asia-Pacific president, during the panel.
Like human employees, AI agents also should be held accountable for actions they make. This means organisations should monitor and have the metrics to measures the performance of their agents, Gupta said.
In addition, companies previously were able to monitor security threats since software design was more static in nature.
This made it more possible to model tools against threats, such as phishing and sequence injection, since these were designed for static threat actors, said Janet George, executive vice president of Mastercard’s AI Center of Excellence.
Agentic AI collapses all of this, creating an environment where agents now have autonomy and can take actions on someone’s behalf, using their credentials, noted George, who also was in the panel discussion.
She added that agents have episodic memory, which can be poisoned by threat actors.
An organisation’s architecture is very different in an agentic world, she stressed.
Cybersecurity teams now have to protect an environment that is dynamic, adaptable, and non-human, she noted.
“This becomes a very profound challenge,” she said.
George urged companies to build security by design and think about how to establish responsible AI at an architectural level.
Every AI agent must have a creator, be authenticated, have an identity, and adhere to the necessary access controls, she said.
Every agent, from its birth and across its entire lifecycle, must be guardrailed, monitored, and journaled, she added.
Trust more crucial as AI goes deeper
This is particularly critical in healthcare, where trust in AI is the foundation for patient safety.
There needs to be rigorous assessment and governance of its use as AI plays a more active role in the Singapore’s healthcare sector, said Rahayu Mahzam, Minister of State for Digital Development and Information and Health, during her opening address at the summit.
Rahayu pointed to local healthcare cluster, SingHealth, which developed its “S.C.O.R.E.” framework to evaluate LLM (large language model) outputs and AI-generated responses in clinical settings. The approach assesses AI models for safety, context and consensus, objectivity, reproducibility, and explainability.
S.C.O.R.E. currently is applied by SingHealth’s clinical teams to evaluate AI tools that facilitate a range of patient care services, including medication enquiry chatbots and AI systems that support specialists during consultations.
It also is used to validate model outputs before deployment and guide the selection of suitable models for different clinical contexts.
The Singapore government has laid out such guidance via various efforts, including the AI in Healthcare Guidelines and Model AI Governance Framework, Rahayu said.
She further underscored the need for practitioners to have the expertise to better apply AI in their field of work.
“AI has the potential to transform what is diagnostically possible and equip clinicians with better tools. But realising that potential depends on how well we bridge two worlds,” she said. “We need clinicians who understand AI well enough to ask better questions of it, identify where it falls short, and ensure it is applied appropriately at the bedside.”
“We need AI researchers who understand clinical care well enough to build tools — not just technically impressive ones — but those that address the realities and fit into how care is delivered,” she added. “It is the combination of clinical expertise with technical capability, applied ethically and responsibly, that turns promising technology into better outcomes for patients.”
A couple of agreements in healthcare were inked at the summit, including a partnership between Singapore General Hospital and A*Star’s Diagnostics Development Hub, which will look to convert AI research initiatives into real-world impact.
These include an AI-powered digital drawing application to identify early memory problems in seniors and a digital twin model of human biology to estimate the risk of chronic kidney disease in patients with type 2 diabetes.
Singapore has the opportunity to reflect on how it can better support an ageing population, Rahayu said.
By 2030, for the first time in its history, the country will have more seniors than children, where one in four Singaporeans will be aged 65 and above, she noted.
“We can use data to detect health patterns before they become disease burdens,” she said. “We can deploy AI to catch what the human eye misses, rather than wait for patients to show up at the clinics.”
From research labs to real-world impact
Beyond healthcare, Singapore wants to drive meaningful adoption of AI through what it calls national missions. These are prioritised for four sectors — advanced manufacturing, healthcare, finance, and connectivity — which collectively account for 40% of the country’s GDP.
The verticals further provide key government drivers, such as data access and regulatory sandboxes, that can facilitate AI innovation, said Josephine Teo, Singapore’s Minister of Digital Development and Information (MDDI) and Minister-in-charge of Cyber Security Agency and Smart Nation, during her keynote address at ATxSummit.
In aviation, for example, the construction of Singapore’s new Terminal 5 will offer the platform for the country to rethink how its air hub operates.
When completed, the terminal at Changi Airport is expected to double passenger handling capacity from 70 million passengers a year to 140 million in the next decade.
Singapore’s Changi currently is the world’s fourth busiest international airport.
Teo said: “How will passengers move from one gate to another? How will baggage be delivered across multiple terminals? How will aircraft landings and takeoffs be sequenced on our runways?”
These challenges need both hardware and software innovations to resolve, she said, adding that a new terminal alone cannot do the job.
A next-generation air traffic management system that prioritises safety, not just volume, also will be crucial, she noted.
It presents opportunities that AI can help with, as it can in Singapore’s maritime hub, which currently houses the world’s largest automated container terminal, she said.
She added that there are rich datasets of complex operations useful for supporting the development of new AI-powered solutions.
Advanced manufacturing, too, offers significant potential for AI to have scale and scope.
Describing Singapore as a “living lab” for the world, Teo noted that developments in physical and embodied AI have great relevance to its manufacturing companies, which already have had to operate efficiently amidst intense competition.
Industry robot density in the country is some five-times the global average and amongst the highest in the world, she said.
“Physical AI can help with simulations for process design [and] better digital twins can improve predictive maintenance, reduce material wastage, and production downtime,” she added.
However, ideas developed in research labs do not always perform well enough in real factory settings, she noted.
To address this, collaborations are needed within the hardware, software, and operational sectors, she said.
Singapore hopes to facilitate this by developing its Punggol Digital District as a testbed for such partnerships, including providing sandboxes for experimentation.
“We will create an integrated data platform, design real-world test scenarios, and provide special testing permits for robot deployment,” Teo said.
Robots, for one, can help workers enhance service delivery to areas that currently are underserved, she said.
Governance crucial even as tech outpaces frameworks
However, while AI has progressed at an accelerated pace, governance frameworks are far from settled, Teo said, as she reiterated Rahayu’s call for trustworthy AI.
With AI becoming more deeply integrated into areas that can affect lives, including healthcare and finance, governance will be even more crucial to ensure AI is deployed responsibly and securely, Teo said.
The risks from autonomous agents span cybersecurity threats and the erosion of trust in information, she said.
“They do not respect national boundaries,” the minister said.
She noted that benefits from AI also may be narrowly distributed, where reduced access and lack of inclusion will disadvantage smaller nations, such as Singapore.
“Internationally-recognised rules and standards will be important, but they will take time to form,” Teo said.
She added that Singapore has taken the first steps in that direction, monitoring how such rules “shape ground realities” and adjusting them as things progress.
She pointed to the government’s Model Governance Framework for Agentic AI, which now has been updated, just months after it was launched in January.
The update includes case studies of real-world agentic deployments by global organisations, including PwC and Workday.
The Singapore government, too, apply the same iterative process to its own use of AI agents, experimenting with the use of sandboxes, Teo said.
