By Mary Li, CEO and Founder, Atlas
Wednesday night, Kem, our CTO, spent more than two hours walking us through the latest release of DSH (DeepSeek Harness). I went in expecting a technical session. I came out with a clearer view of Atlas — and a different understanding of “AI Native” and “capability.”
What Atlas is really building
One idea in DSH stayed with me: almost everything can become a plugin. Models change. Tools change. Skills can be installed. Even the agent loop can be replaced. It made me ask a simple question: if intelligence becomes increasingly available and interchangeable, what do agents still need that they cannot create themselves?
For LCC distribution, I believe the answer is data, fulfillment and payments.
Data is not just fares and availability. It is reliable airline content, together with the data layer customers and agents need to understand and use it.
Fulfillment means getting the booking done and looking after it afterwards. Ticketing, ancillaries, changes, refunds and exceptions. Search finds an answer. Fulfillment takes responsibility for the outcome.
Payments is how money moves safely through the system: acquiring, airline settlement, risk control and reconciliation.
None of these came from a model. They came from years of integrations, operations, failures, exceptions and learning.
This makes our direction clearer to me. Atlas should not try to build everything an agent needs. We should make these three capabilities easy for every agent, seller and partner to use.
AI Native keeps moving
DSH also changed my understanding of “AI Native.”
The center of gravity is moving away from the model itself. What matters increasingly is the system around it: context, memory, tools, permissions, environment, evaluation and the ability to keep working toward an outcome.
Codex, OpenClaw and DSH are interesting to me for this reason. Intelligence is becoming modular. And the change is happening so quickly that our definition of AI Native can become outdated within weeks.
So I have changed my own definition.
Being AI Native is not about how much AI a company uses. It is about how much of the organization is willing to redesign itself around what AI can now do.
That changes management too. If agents can work out more of the steps themselves, managers should spend less time defining every step. The job moves toward setting the goal, defining boundaries, preserving what has been learned and measuring the outcome.
This sounds like an AI idea. In many ways, it is simply good management becoming more important.
What is a capability now?
This also made me rethink a word we use all the time at Atlas: capability.
I used to describe capability as skills plus process, data, experience and teamwork, producing an outcome customers can rely on. I still believe that. But the moat is moving. If skills can be installed, models switched and tools replaced, more of what we once considered expertise will become available to everyone.
What is harder to copy is what an organization has learned.
Every failed booking we investigated, every unusual airline rule we discovered, every refund exception and payment problem we solved has added something to Atlas.
The important question now is whether we can turn that accumulated experience into organizational intelligence — something our people and our agents can both use, and keep improving.
The same question applies to each of us.“What skills do I have?” is becoming less interesting.
A better question is: what have I learned that makes the whole system better?
What judgment have I contributed? What knowledge have I made reusable? What can the organization now do more reliably because of what I learned? Perhaps that is also where people matter most in an AI Native organization.
Technology will keep moving, probably faster than our definitions can keep up. Good. It means we have to keep learning too.
Onward and upward.
