At the recently concluded AMD AI Developer Day 2026, Kai-Fu Lee, CEO of 01.AI, engaged in a dialogue with Lisa Su, Chair and CEO of Advanced Micro Devices (AMD), on the topic of "New Paradigms for AI Agents." Lee pointed out that generative AI is advancing towards the era of agents. This year, AI programming capabilities have crossed a critical threshold, and multi-agent architectures are increasingly discussed for their ability to break through the limitations of single agents.
Regarding the "AI transformation" that concerns most corporate CEOs, Lee offered his own advice: do not rely solely on your CIO. Enterprise AI transformation is unequivocally a top-down leadership project requiring a fundamental shift in mindset from business leaders. He urged companies to stop superficial "showcase AI" and begin building structural engines that can genuinely integrate with core business operations.
He also introduced an operational framework to break down traditional boundaries: the concept of the DRI (Directly Responsible Individual). He further noted that the working methods of many engineers today will change—they will no longer focus solely on systems but will be accountable for outcomes.
"This year, AI has undergone two key changes, crossing the 'autonomous execution' threshold." Discussing observations on AI industry development, Lee noted that a year ago, AI could only assist in writing code, functions, etc. Now, it can deliver an entire set of features end-to-end. This may sound like an incremental improvement, but it is not. All actions of an agent in the digital world ultimately manifest at the code level. Once AI's coding ability crosses that threshold, autonomous agents truly become a reality.
Furthermore, a more significant change is the realization that a single agent has inherent capability limits. Regardless of model parameter size, relying solely on the reasoning power of a single Agent will eventually encounter bottlenecks when facing real-world complex problems. Multi-agent architectures break this ceiling for the first time. Different agents responsible for planning, evaluation, research, and execution begin to collaborate, debate, and iterate on each other's results.
In Lee's view, this closely resembles the "Medici Effect": when experts from different fields are brought into the same room, the final outcome far exceeds the simple sum of any single individual's capabilities. Five hundred years ago, during the Renaissance, humans discovered this principle. It is only in the 21st century that we have brought this mechanism into the AI world for the first time.
From a technical perspective, this means we are gradually moving away from the past model of "trying to use one model to do everything." Future AI will not be a solo performance by a "super brain" but more like a symphony performed by different intelligent systems working in concert. Based on this trend, we are deploying specialized multi-agent systems and gradually moving towards a stage of "Heterogeneous Intelligence." Different types of models and algorithms will be combined to use collective intelligence to solve more complex problems.
The most pressing question in 2024 was: "Can AI complete a task?" In 2025, the question evolved to: "Can AI complete an entire workflow?" In 2026, the core question has advanced to: "Can AI replace an entire corporate functional department?" Using a modern HR (Human Resources) department as an example, he pointed out that when a recruitment Agent and a performance Agent are linked, the system can automatically adjust front-end talent screening criteria based on actual performance data after an employee joins. From resume screening and interviews to new employee onboarding, and automated monthly and quarterly performance tracking, these multi-agent systems will continuously operate and upgrade around unified HR data. As this capability expands, it will eventually evolve into an interconnected enterprise multi-agent collaboration network covering different departments such as HR, R&D, Product, Sales, and Marketing.
This architecture is also driving the emergence of the "One-Person Company" trend. Leveraging a modular multi-agent framework, individual developers or domain experts now have the ability, like a "chief architect," to rapidly launch a highly automated company.
"Under the new agent-driven paradigm, we have essentially crossed the 'autonomous execution' threshold. AI is shifting from the past passive 'Prompt-and-Response' mode to an active 'Goal-and-Execution' mode," Lee stated. In his view, in the future, you won't just give AI a prompt; you will give it an organizational goal. Subsequently, the agents will autonomously complete coordination, execution, evaluation, optimization, and form a complete closed loop.
"Stop Showcase AI, Transformation Cannot Rely Solely on the CIO" The shift in the AI paradigm is also creating the largest commercial opportunity in the current AI field: industrial-grade AI transformation. The real economic value in the new era will not come from AI systems that merely "answer questions," but from autonomous multi-agent infrastructure capable of truly executing corporate objectives.
Lee pointed out that in the agent era, truly "autonomous enterprises" will be born. They will be driven by cross-departmental, multi-level collaborative agent networks. The next phase of industrial AI transformation will revolve around two core issues simultaneously: data sovereignty and clear, verifiable ROI (Return on Investment).
For developers, the greatest opportunity lies in building AIs that previously required an entire team to complete but can now independently deliver commercial outcomes. The role of AI is no longer "an AI tool to help marketers improve efficiency," but an AI Agent that can genuinely assume marketing functions; not "an AI tool assisting financial analysts," but an AI Agent capable of providing automated financial analysis.
For the vast majority of corporate CEOs, Lee also offered his "enterprise AI transformation" advice—do not rely solely on your CIO.
Lee candidly shared that in his discussions with numerous CEOs about AI transformation, he found that almost every company is currently choosing to deploy AI in low-risk but low-value scenarios. Examples include meeting minutes, HR employee Q&A chatbots, internal enterprise search, etc. However, these are merely superficial applications.
"I tell CEOs bluntly, do not just listen to your CIOs. Typical CIOs focus on system stability, software security, and error-free operation. In this round of AI transformation that deeply affects the core business lifeline of enterprises, they may instead become the old guard hindering evolution," he stated. In his view, because the CIO's responsibility is essentially managing software operations, not redefining the company, CIOs excel at deploying AI safely but are not adept at driving genuine organizational-level change.
"Most AI transformations driven bottom-up by IT departments will ultimately fail," he believes. The traditional CIO role will not disappear, but its importance will be significantly diminished because AI is not just another new software tool. Enterprise AI transformation is absolutely a top leadership project requiring a fundamental mindset shift from business leaders.
"What truly changes a company's operating results are often the core business processes that directly impact the Profit and Loss (P&L) statement. And these areas are precisely the operational functions where many executives are most reluctant to let AI intervene: revenue, profit, fraud prevention, dynamic pricing, supply chain, product launch speed, and core innovation capabilities," Lee stated directly. "Forward-looking CEOs are recalibrating how their companies operate, how organizations should change, and how leadership styles should adapt."
In his view, anyone involved in business development should think in the same way. Stop superficial "showcase AI" and start building structural engines that can truly penetrate the substance of the business.
"The DRI Model Redefines Tech Professionals; Everyone Should Think Like a CEO" Today, a single individual with the right tools and sufficient computing power can accomplish work that required an entire team just a few years ago.
Observing those who are truly pioneering and developing in this way, Lee recently noted that, constrained by past training, most developers are accustomed to thinking about ownership issues at the code level. For example, one person is responsible for a code repository and PR (Pull Request) on GitHub, another for on-call rotation, and another for a specific service. This responsibility boundary essentially only covers the parts you can directly control via your keyboard. Now, more and more coding work is beginning to be taken over by AI agents.
Addressing this change, Lee proposed an operational framework to break down responsibility boundaries: the concept of the DRI (Directly Responsible Individual). He predicted that the DRI model will become the core organizational structure for AI-native companies.
A DRI is an individual who assumes end-to-end responsibility for a cross-functional outcome. It is not a job title but a very clear accountability mechanism. Like the sole designated on-call engineer in a system runbook: the final result and its business impact are the responsibility of the DRI.
In software engineering, the main bottleneck in delivering a product is rarely the code itself. It is the ambiguity of ownership. Dispersed responsibility, stalled pull requests, and deviated roadmaps often stem from many people being responsible for a single link in a project management spreadsheet, with no one truly accountable for the final outcome. In the DRI model, a human DRI sits at the center of the entire agent system, coordinating with a specialized cluster of different Agents (research, execution, compliance, monitoring, etc.). The DRI does not spend time and energy on specific execution but is responsible for overall orchestration, key decisions, and accountability for the final output contract. Meanwhile, real-time data streams will gradually replace traditional reporting systems, and business operations will increasingly revolve around specific, quantifiable results.
"I believe that many of the capabilities possessed by an excellent engineer are highly consistent with those needed to be an excellent DRI," Lee stated. When you write technical specifications, you are essentially trying to define quantifiable business outcomes. When you monitor a system and configure automatic alerts, you are establishing mechanisms to measure results. When you proactively debug and troubleshoot issues at 2 AM instead of waiting for someone to notify you, you are demonstrating the core sense of ownership central to the DRI model.
In his view, choosing the DRI model also means you must redefine "what constitutes personal success." In the agent era, the value of an excellent engineer is no longer measured solely by "how much code was written." This also means the working methods of many engineers today will change. You no longer just focus on the system; you are responsible for the outcome. Excellent engineers typically place great importance on monitoring systems, making services highly observable. The DRI extends this technical rigor to the business results they own.
"If you are a DRI responsible for growth, you don't just monitor API latency; you also monitor user activation rates, conversion funnels, and impact on revenue. You are responsible end-to-end for the complete outcome," Lee illustrated. "You have decision-making authority, not just advisory power. Engineers are usually good at analysis, but they often hand over the output to product managers or executives to make choices. A DRI needs to close the loop themselves: you analyze, you decide, and you are responsible for whatever happens next. It might feel uncomfortable at first, but you'll get used to it quickly."
Comments