The Agentic AI Opportunity for the Caribbean: Build for the Problems We Actually Have

The global race in artificial intelligence has largely been defined by scale: larger models, more data, more compute and billions of dollars of capital expenditure.

But for the Caribbean, Africa and other resource-constrained and distributed markets, the more important question may be different:

What can we build when intelligence becomes dramatically cheaper, but capital, compute, data and specialized technical talent remain constrained?

That question sat at the center of a recent Future Caribbean fireside conversation with Lucio Dery, Research Scientist at Google DeepMind, hosted alongside Roniesia Gittens of Google DeepMind.

Dery grew up in Ghana before studying physics and computer science at Stanford, working at Facebook AI Research, completing his PhD at Carnegie Mellon and ultimately joining Google DeepMind. His research has focused heavily on resource-efficient transfer learning, continual learning and increasingly the intersection between continual learning and Agentic Systems.

His perspective is particularly relevant to Future Caribbean because it challenges one of the assumptions driving much of the current Agentic AI conversation: that everyone should be trying to reproduce the systems being built by the world's largest technology companies.

For distributed markets, the opportunity may be precisely the opposite.

Intelligence Is Becoming Cheaper

Dery described one of the most consequential shifts created by today's Agentic Systems: access to certain forms of intelligence and technical capability is becoming significantly less expensive.

Historically, building even relatively narrow digital infrastructure could require teams with expertise spanning software engineering, databases, machine learning and other technical disciplines. For countries where capital and specialized talent are constrained, that creates an enormous barrier to deployment.

Agentic Systems begin to change the economics.

With the right specifications and architecture, Dery argued, people without extensive programming expertise can increasingly get working prototypes off the ground. A project that may previously have required six or seven engineers might begin with one highly capable technical builder working alongside increasingly powerful models.

That matters enormously for the Caribbean.

The opportunity is not merely to automate existing work. It is to make previously uneconomic problems economically possible to solve.

What Exactly Is an Agentic System?

Traditional software is built around relatively rigid logic. Developers determine what happens when particular inputs arrive, which systems should be called and what actions follow.

Agentic Systems introduce an intelligent decision-making layer.

An LLM can increasingly determine which tools to use, what information to retrieve, when to retry an action and how to respond to changing circumstances. The resulting system combines the underlying model with the surrounding tools, information and architecture required to accomplish a task.

This becomes even more powerful when combined with continual learning.

Today's foundation models are largely trained through enormous upfront investments and then deployed. In that sense, Dery explained, they are effectively "frozen in time." They do not inherently continue learning from everything that happens after deployment.

Continual learning asks a different question: How do we create systems that become better as they operate?

For an Agentic System acting as a long-term assistant, healthcare system, financial infrastructure layer or operating system, that becomes critical. The system must accumulate useful information, remember previous interactions and adapt its behavior over time.

The Caribbean's Constraints May Point Toward Different AI

Dery's own interest in efficiency emerged partly from growing up in Ghana, where access to abundant compute and resources could not be assumed.

The prevailing formula for improving frontier models has been to make them larger: more parameters, more data, more compute and more capital. But that is not necessarily the architecture that makes the most sense everywhere.

For regions operating under different constraints, efficiency itself becomes an innovation problem.

This is especially relevant for edge AI. Edge environments introduce hard limitations around memory, electricity and internet connectivity. Instead of assuming constant access to enormous models running inside hyperscale data centers, developers may need hybrid architectures: smaller models operating locally, larger models accessed remotely when connectivity permits, and carefully designed systems deciding what needs to happen where.

For island states, rural communities, disaster environments, ocean systems and other infrastructure-constrained settings, these are not peripheral considerations.

They are fundamental design requirements.

Don't Start by Trying to Connect Everything

The region is fragmented across countries, currencies, regulators, legal systems, institutions and markets. Agentic AI appears capable of becoming the connective tissue between them.

Dery's advice was more disciplined. Before attempting to glue fragmented systems together, strengthen the individual pieces.

There may already be substantial opportunities within healthcare, finance, logistics, agriculture, government services and other individual domains where Agentic Systems can reduce the cost of expertise and make previously difficult technology deployments viable.

Trying to connect everything immediately introduces another problem: integration itself is expensive.

And the bottleneck is often not technological. It is human.

Administrative processes, institutional approvals, access to data, regulation and adoption can become larger barriers than the intelligence of the underlying model. Dery pointed to electronic health records as one example: the technology may exist to digitize and intelligently structure records, but obtaining access to the information and navigating administrative processes can still stop deployment.

Agentic AI can reduce technical bottlenecks. It cannot automatically eliminate institutional ones.

Modularity Is Powerful — But It Comes With a Cost

Many emerging Agentic Systems are being designed modularly: specialized agents handle different tasks while orchestration layers coordinate them.

Dividing a system into specialized components can improve efficiency and scalability. At the same time, information becomes distributed across different modules, creating new coordination and search problems.

Imagine one agent responsible for travel, another for healthcare and another for financial activity. Each may perform extremely well independently. But the moment a user's objective touches all three domains, the system must reconstruct context across those separate components.

The lesson is not to avoid modularity. It is to recognize that coordination is itself an intelligence problem.

When a Future Caribbean builder asked how to prove that a multi-agent architecture was demonstrating "genuine reasoning" rather than merely orchestrating narrow models, Dery challenged the premise. Selecting the right specialized solution for the right problem may itself constitute valuable reasoning. What ultimately matters is whether the system reliably performs the task it was designed to accomplish.

As he put it, builders should "look at the numbers": performance and accuracy matter more than proving an abstract distinction between general reasoning and effective orchestration.

Data Scarcity Requires a Different Playbook

Many important regional problems do not come with billions of clean training examples. Data may be sparse, fragmented, intermittent or simply unavailable.

That does not make machine learning impossible. It changes the methodology.

Dery highlighted transfer learning as one particularly important approach: begin with a system exposed to a larger pool of related data and then adapt it using the smaller amount of domain-specific data available. As additional local data arrives, specialized models can continue being retrained and adapted.

In data-scarce environments, developers may need to impose more structure and constraints themselves. Transfer learning, data augmentation, generative approaches and carefully designed domain knowledge become considerably more important.

The Models Are Global. Their Training Context Is Not.

Today's leading models are often described as general-purpose systems. But Dery cautioned that, when deployed in practice, many remain highly optimized around Western — particularly U.S. — environments.

Builders in the Caribbean cannot assume a system developed elsewhere will map perfectly onto local institutions, languages, workflows, regulations or realities.

This creates an enormous opportunity. The competitive advantage of Caribbean builders may not be having more compute than Silicon Valley. It may be having something Silicon Valley does not: context.

Understanding how Caribbean healthcare actually operates. Regional payments and settlement. Disaster response across islands. Informal economic activity. Transportation between fragmented jurisdictions. The realities of Caribbean agriculture, tourism, energy and ocean systems.

Agentic AI lowers the cost of building the intelligence layer. Local builders supply the domain understanding required to make that intelligence useful.

From Extraction Infrastructure to Settlement Infrastructure

Many Caribbean and Global South systems were historically constructed around extraction: moving resources, capital and value outward.

What would it mean to build the opposite? Infrastructure designed for settlement, accumulation and growth.

Financial rails are one example. Investment infrastructure is another. But the idea extends further: digital systems that allow fragmented economies to coordinate, transact, accumulate capital, share intelligence and participate more effectively in larger markets.

The path may not begin with one enormous operating system attempting to solve everything. It may begin with highly focused systems solving real problems extraordinarily well — healthcare, payments, logistics, disaster coordination, ocean intelligence, investment, government services — and gradually creating the interfaces that allow those systems to work together.

Strengthen the nodes. Then build the network.

Build Under Constraint

The region does not need to wait until it possesses the compute, capital or engineering density of the world's largest technology centers.

The economics of building are changing. Problems that were previously too expensive to address may suddenly become viable. Expertise that was previously inaccessible can increasingly be augmented by models. Small technical teams can attempt projects that would have required substantially larger organizations only a few years ago.

But cheaper intelligence does not eliminate the need for judgment, domain expertise, good architecture, safety or institutional cooperation. It makes those things more important.

Dery closed with a message directly relevant to the builders working across Future Caribbean: build for local problems and bring your own perspective to them. As intelligence becomes cheaper, the human processes surrounding these systems — how they are designed, governed, integrated and deployed — become increasingly important.

The Caribbean does not need to reproduce somebody else's AI infrastructure.

It can design systems around the realities of distributed markets, constrained resources, fragmented jurisdictions and the problems its people actually need solved.

And if those systems work here, their relevance may extend far beyond the Caribbean.