The Next Wave of AI: from Generation to Agency

By Jacques Bughin

If you’re quietly patting yourself on the back for finally getting a grip on AI and its impact on your organization, here’s a reality check. It turns out there is more to AI than GenAI. Agentic AI is coming our way – and this time, it’s REALLY big!

The last two years have been dominated by generative AI, LLMs that produce text, code, and images at unprecedented speed. But the frontier has shifted. A new wave is forming, one that is less about generating content and far more about taking action. This wave is agentic AI,¹ systems that plan, decide, execute, coordinate with other agents, and interface with real-world tools, software, or machines. And, unlike the previous transitions, this one fundamentally reshapes entire industries, labor markets, and the competitive landscape of AI firms.

Unlike the previous transitions, this one fundamentally reshapes entire industries, labor markets, and the competitive landscape of AI firms.

To understand why, one must look beyond the hype and examine what the emerging players are actually doing. Across the world, we see the foundational pieces of an agentic economy being assembled. Some companies—Moveworks,² ServiceNow, OpenAI, Anthropic, CrewAI, LangGraph—are building the orchestration and multi-agent fabric. Others—Alibaba DingTalk, Tencent, ByteDance, Baidu—are deploying agents at societal scale. In Europe, Siemens and ABB are embedding agents inside factories, robots, and supply chains. Yet, while the surface impression is progress, the deeper truth is that the global market is still mono-agent, doing tool calls rather than cooperation. True multi-agent systems, agents coordinating as teams, are only in their infancy.

However, the direction of travel is now known; we are entering a world where most workflow, coordination, planning, and even knowledge work will be executed by agentic systems, not by generative models. And this shift will be bigger and more transformative than the GenAI wave for three reasons:

1. agency automates tasks, not content;
2. agency scales labor;
3. agency restructures firms, workflows, and entire industries.

GenAI revolutionized output creation. Agentic AI is the wave we need to revolutionize the entire concept of work.

1. The emerging agentic market: still mono-agent, but crystallizing fast

Although the industry uses the language of “multi-agent AI,” today’s systems are, bluntly, one-agent wrappers around LLMs. Moveworks, the leader in enterprise AI service management, provides a single enterprise assistant that resolves tickets, completes HR workflows, and updates internal systems. It behaves like a highly competent internal employee: it resets passwords, rewrites policies, updates CRM fields, books travel, and links to Jira or Workday. But all this flows from a single agent orchestrating many tool calls. It is not yet coordinating with other autonomous agents; rather, it is acting as a “meta-employee” for the enterprise.

The same is true for Microsoft’s Autogen-based internal systems, and for Google’s Gemini Code Assist. They are not yet multi-agent societies; they are intelligent single execution loops with planning sequences.

Only a few players push into real multi-agent autonomy. CrewAI is an open-source Python library which allows the orchestration of multiple AI agents as a real project team. Instead of settling for a single generalist assistant, one can create a squad of specialized AIs, each with a role, a mission, and the ability to communicate with their colleagues. The agents are powered by LLMs such as GPT-4o or Claude. Each agent acts within their field, collaborates with others, and contributes to advancing the mission. Everything is coordinated by a manager agent who orchestrates the team. What makes CrewAI so powerful is its role-based agents, its shared contextual memory system, and its ability to handle complex inter-agent conversations. LangGraph is another case in point as an orchestration framework designed for building multi-agent AI systems with large language models. It allows developers to create complex, dynamic workflows as graph structures where multiple AI agents interact, collaborate, and maintain context and state over long-running tasks. LangGraph excels in managing multi-agent communications with fine-grained control over application flow, enabling reliable, customizable, and scalable agentic AI applications, including conversational agents, task automation, and decision-making systems.

But even here, most use is experimental. Agents cooperate to write reports, run simulations, or perform market analysis, not to autonomously run supply chains or operate financial systems. China is the exception. Its industrial platforms—DingTalk, Tencent’s scenario platforms, Baidu’s Apollo, ByteDance’s commerce systems—use multi-agent structures, because the underlying digital ecosystems are unified. DingTalk agents negotiate task assignments and approval flows. ByteDance’s HiAgent allows pricing, logistics, advertising, and inventory agents to coordinate asynchronously. Baidu Apollo’s self-driving system is multi-agent by necessity, allowing vehicles to learn from collective driving experiences and scenario data. This distributed multi-agent structure enables scalability and fleet-level optimization, supporting real-time scenario simulation, validation, and model updates that enhance safety and performance. The necessity for this multi-agent approach stems from the complexity and safety-critical demands of autonomous driving. No single monolithic model can efficiently and reliably handle the wide range of subtasks required in diverse driving environments. Instead, modular agents specializing in perception, mapping, prediction, and planning enable parallel processing, robustness, and modular upgrades. Real-time interaction of these agents ensures continuous adaptation to changing conditions, while fleet-wide coordination facilitates system improvements on a large scale.

2. Why agentic AI is the next wave

(and bigger than generative AI)

To understand why agentic AI will overshadow generative AI, we must look at what it fundamentally changes. Generative AI produces content. That is powerful; coding assistants like Cursor or ChatGPT can generate boilerplate, transform legacy systems, and help developers. But content generation has natural limits: content is the output of tasks, not the tasks themselves.

Agentic AI flips this relationship. It automates the task, not the output. Instead of generating an email, the agent reads the email, opens Salesforce, retrieves context, drafts a response, updates the opportunity, books a meeting, and files a ticket. Instead of summarizing policy documents, the agent updates compliance workflows, sends approval requests, writes the audit trail, and coordinates with five other internal systems.

An agent is not a “smart model”; it is a worker. And when workers scale, so does value creation. This shift has three systemic implications:

First, agentic AI attacks the coordination costs of the firm, the deepest cost structure identified by Coase. If an agent can schedule meetings, allocate resources, file claims, run ETL pipelines, reconcile invoices, and coordinate inventory, the firm transforms from a hierarchy of labor into a network of autonomous processes. Productivity is no longer linked to headcount.

Second, agentic AI enables stacked multipliers: one agent helps sales; ten agents run an entire sales pipeline; fifty agents automate a supply chain. Generative models do not scale this way.

Third, agentic AI displaces GenAI-native companies because generation is commoditized. Once agents automate tasks directly, the value shifts from content to execution. Cursor helped developers write code; agentic systems automate the entire issue lifecycle: triage → fix → test → deploy → notify stakeholders. The more agentic systems mature, the weaker the standalone GenAI-only products become.

3. Why agentic AI will reshape employment

(more than GenAI ever could)

Predictive AI affected forecasters and analysts. GenAI affected writers, coders, and creatives. Agentic AI affects everyone, because it automates the workflows that make up jobs.

Moreover, multi-agent systems will automate coordination, the highest-level human activity in firms.

Moveworks and ServiceNow already demonstrate that a single agent can absorb 40–70 percent of IT and HR tickets. In major companies, this is the equivalent of replacing dozens of support staff. ByteDance HiAgent coordinates advertising, logistics, and customer support, reducing labor requirements across multiple domains simultaneously. DingTalk agents in China automate HR, finance, and purchasing workflows for millions of SMEs. Unlike GenAI, which “augments,” agentic AI executes. It can read emails, log into systems, reason over multi-step workflows, call APIs and make decisions

Moreover, multi-agent systems will automate coordination, the highest-level human activity in firms: resource allocation, scheduling, negotiation, prioritization. This is why the shift is more profound than the move to GenAI. GenAI replaced creation, but agentic AI replaces coordination, which is what managers, administrators, and entire corporate functions are paid to do.

4. Agentic AI may oblige GenAI-only startups to reinvent

The GenAI SaaS wave (2020–3) produced an explosion of startups offering “smart content.” But the economics of agentic AI destroy that value proposition. A GenAI-only product generates a document, a query, or a piece of code. An agentic AI system reads the requirement, executes the task, interacts with systems, and completes the process.

Cursor is already facing this reality. Although it is a brilliant coding assistant, agentic systems like Devin or GPT-based Code Agents can automate entire tickets end to end, making a coding editor assistant insufficient. Jasper and Copy.ai have declined sharply in usage because marketing agents can now plan campaigns, test variants, analyze CTR, adjust budgets, and post on social media, not just generate copy. The more agentic AI improves, the more GenAI-only tools lose relevance. Why use a coding assistant when an agent can build, test, deploy, and monitor features? Why use a customer-service chatbot when an agent can resolve the case?

GenAI tools focused on “generation” become components, not products. Agentic AI is not a new product category; it is a platform shift that absorbs generation entirely.

As an example, consider Moveworks. It represents the first generation of enterprise agentic platforms, a single agent with deep enterprise integration and thousands of pretrained workflows. Its competitive strength lies in the density of integrations, not the intelligence of the model. It is a mono-agent that behaves like an entire tier-1 support team. This is why ServiceNow acquired Moveworks: it fits into a broader agentic vision where each enterprise function gets an autonomous system.

CrewAI and LangGraph represent the second generation—multi-agent orchestration frameworks, where different agents assume different roles, negotiate tasks, and pass control. These frameworks are early, messy, and experimental, but they are the seeds of a future where enterprises run dozens or hundreds of cooperating agents across departments. In China, DingTalk and ByteDance are already moving toward multi-agent ecosystems with specialized agents that cooperate across logistics, finance, inventory, marketing, and HR. In many ways, China is executing the true multi-agent vision earlier, because its digital ecosystems are unified.

Conclusion: The age of agency will restructure the economy; be ready for it

The next wave of AI is not about models but about actions. It is not about intelligence but about coordination. It is not about content but about workflows. Agentic AI will reshape firms, collapse coordination costs, create new digital labor forces, disrupt GenAI-only companies, and permanently alter labor markets.

Mono-agent systems will dominate in the short term, but multi-agent cooperation will define the long-term landscape. China is ahead in deployment, the US in frameworks, and Europe in industrial integration, if it can lower its cost of deployment. Agentic AI is not the next step after GenAI. It is likely a replacement for it. The era of autonomous work besides simple robotics has begun.

Managers must shift from supervising workflows to owning outcomes, because agentic AI automates the coordination tasks that once defined managerial work. Their role becomes that of system architect, not task allocator, designing which workflows agents execute, setting guardrails, and auditing AI decisions. They must manage constraints, not steps: accuracy thresholds, compliance logic, escalation paths, and risk boundaries. Data stewardship becomes central, since agentic AI’s performance depends on clean data flows, standardized processes, and interoperable systems. Metrics move from micro-monitoring humans to macro-monitoring system productivity, error vectors, and escalation patterns. Managers must redeploy humans into high-judgment roles: exception handling, negotiation, creativity, and cross-functional sense-making. They must also master AI risk management through audits, drift monitoring, red-teaming, and scenario testing.

Ultimately, managers evolve into hybrid orchestrators of humans and agents, responsible for strategic alignment, workflow design, constraint definition, and organizational learning. The quantity of managerial labor declines, but the strategic intensity of what remains increases sharply.

About the Author

Jacques Bughin is the CEO of MachaonAdvisory and a former professor of Management. He retired from McKinsey as a senior partner and director of the McKinsey Global Institute. He advises Antler and Fortino Capital, two major VC /PE firms, and serves on the board of several companies.