Agentic AI could usher in a new era of AI capabilities

Imagine you’re a mid-level manager at a tech company. One day, upper management announces that it’s assigning you a new intern. His name is Paul, and he’s supposed to help with website development.

Like many developers on your team, Paul works remotely. You put him right to work on a new site for your firm, giving him a list of tasks and specific guidelines.

Photo credit: Shutterstock

To your surprise, he comes back at the end of the day with a fully working site, complete with individual pages and visuals. It’s mostly bug-free, too.

Over the next few weeks, you notice that even without prompting, Paul constantly updates the site to address any issue that arises.

“What a perfect intern,” you think to yourself.

Here’s the catch: Paul is not a human. He’s an “agentic AI” – a new type of AI model that could be upon us very soon.

Taking initiative

Essentially, agentic AI is a model that can automatically make its own decisions to fulfill an objective.

Agentic AI isn’t like the general AI solutions we currently have – which include popular large language models like ChatGPT and Gemini – that mostly require constant prompting to stay on the right track.

“The promise of agentic AI is about getting deterministic results through an indeterministic system. In short, you tell it the end goal and it can get there on its own,” says Christopher Cai, head of product and engineering at software development company AngelHack Developer Labs.

Christopher Cai, head of product and engineering at AngelHack Developer Labs / Photo credit: Tribe

Agentic AI can work with a prompt like, “Plan me a fun holiday to Thailand,” instead of needing explicit instructions like, “Compile a list of fun tourist attractions, restaurants, hotels, and transport options for a vacation in Thailand.”

In a business setting, an agentic AI model can be used to invoice a company’s clients. All it would need are the details of the clients and clear guidelines on when the invoices need to be issued.

Once these are provided, the agentic AI model generates the invoices on its own. It then ships them off as soon as the conditions – such as fulfillment of services – are met.

While today’s AI solutions may be able to generate invoices for you, you’d still have to tell them – step by step – which details to fill out.

“With agentic AI, you wouldn’t need any human intervention to tell it what to do at every possible stage or scenario,” Cai points out. “It can think of what to do on its own within the parameters you set.”

Self-directed learning

The idea of an agentic AI is exciting, but like every other AI type, it’s faced with the question of reliability. It’s no surprise, given that even some of the biggest and most advanced models today, such as GPT-4, experience hallucinations at varying rates.

According to Cai, it will be crucial to have robust policy and governance to ensure that the data accessed by agentic AI models is used properly and only for relevant tasks.

“It’s less of an AI-only problem and more of how to engineer the checks and balances in the right areas,” he explains.

It’s also important to make sure that agentic AI models have enough high-quality data to train on. This way, they are prepared for a wide range of possible scenarios.

Photo credit: Shutterstock

Additionally, Cai highlights the concept of having an internal feedback system within or between agentic AI models. Instead of relying on manually tweaking the algorithm to improve a model, AI would be able to “check its own work,” so to speak.

“For example, the GPT-4 models are more accurate because they can process the intent of the prompt, take the response generated, and then feed it through the model again to see if it makes sense,” he explains.

A future with agentic AI could mean having different agents interacting with one another within the same system. For example, one agent may do the coding while another may perform quality checks to ensure that the code is correct. Each AI agent specializes in a particular domain, exchanging results with other agents to create a feedback loop that improves each model over time.

Integrating AI into the workplace

Even if we were to successfully develop such capabilities in agents, implementing them might take a while. With so much data floating around in any given organization, figuring out how to effectively roll out tech can be challenging.

“Think about how many SaaS applications we have,” says Cai. “Creating an agentic AI model that can integrate with and pull accurate data from all of them could be a nightmare.”

This is particularly important “in sensitive areas like healthcare or finance,” he adds.

Photo credit: Shutterstock

There are no two ways around it. Advocates of agentic AI will need to educate people to show that the models work and can be trusted.

Moreover, adoption will have to move at a gradual pace. Implementing domain-specific regulations, which can be modeled after existing policies, could be a good start.

For example, an agentic AI model used for HR functions may deal with sensitive personal details, like those in performance reviews. Applying existing standards to AI would mean it could only access these pieces of information if the task is given by someone who already has access to them, like a manager.

“It’s impossible to catch every single thing early in development. But we should start with policies that are already in place and translate them into guardrails for the AI,” says Cai.

Timeline for adoption

Cai thinks agentic AI presents thrilling possibilities, and widespread adoption could be a reality very soon – especially with major players like Nvidia now focusing on it. The chipmaker recently launched AI Blueprints to help developers build and deploy their own custom AI agents.

With these foundational layers in place, it’s just a matter of hard work and trial and error to get on the right track.

“It’s similar to the SaaS wave, which took a few iterations before we saw the big companies coming up with solutions and proper use cases,” Cai recalls.

“Agentic AI technology will go through the same experimentation for various use cases, and once that’s done, we’ll start to see it as a core part of working, he adds.

Backed by the Government of Singapore, Tribe is a leading startup accelerator focused on driving innovation in the global AI ecosystem. It runs AngelHack DevLabs, which helps firms build innovative digital products and bolster their tech team. Visit its site to find out more.

FIND OUT MORE

This content was produced by Tech in Asia Studios, which connects brands with Asia’s tech community. Learn more about partnering with Tech in Asia Studios.