How Far for Stage 2? Persistent Agentic Systems and the Real Meaning of AI

The wrong question is still being asked.

Most public and organisational discussion about artificial intelligence remains trapped in a theatrical argument about whether machines are “conscious,” whether they “think,” when they will be or whether they are merely “stochastic parrots.”

That debate, beyond the hype, may be philosophically interesting, but it is strategically distracting. The more urgent question for business, society, education, government, and the future of work is simpler:

How far are we from persistent agentic systems becoming normal, reliable, trusted, and operational?

The answer is uncomfortable because it removes the luxury of delay (which never really existed with any credibility as delay is legacy and that equates to strategic drift and demise).

Put another way, we are not waiting for artificial intelligence to arrive. We are already living through the early infrastructure build-out of persistent agentic systems. The issue is not arrival. The issue is maturity.

Stage 1 AI was conversational. It answered, summarised, drafted, translated, coded, explained, and advised. It was powerful, but largely episodic. You opened a window, asked a question, received an output, and moved on.

Stage 2 AI is different. It is persistent. It remembers. It acts. It connects to tools. It monitors. It schedules. It searches. It drafts. It updates. It learns user preferences over time. It begins to operate across email, calendar, documents, code, CRM, finance systems, workflows, knowledge bases, browsers, and digital workspaces.

That is not “chatbot improvement.” That is the emergence of an operating layer around human work.

How Far for Stage 2? Persistent Agentic Systems and the Real Meaning of AI

The basic version is already here in 2026. Memory across conversations, scheduled monitoring, tool use, browser use, computer-use agents, coding agents, workflow agents, long-term memory services, and enterprise AI orchestration are all now visible parts of the technology stack. OpenAI has continued to develop memory systems that keep context fresh and relevant across conversations. Google Cloud has introduced Memory Bank capabilities for agents to store and retrieve long-term memory across sessions. Anthropic’s Claude supports computer-use-style agents able to interact with digital interfaces, files, browsers, and screen workflows.

The direction of travel is clear. AI is moving from answer machine to action system.

That shift matters because the future of work will not be determined by who has access to AI. It will be determined by who redesigns work around AI.

Microsoft’s Work Trend Index describes the rise of the “Frontier Firm”: organisations structured around on-demand intelligence and hybrid teams of humans plus agents. Its 2025 report drew on survey data from 31,000 workers across 31 markets, LinkedIn labour-market trends, and Microsoft 365 productivity signals. The conclusion was not subtle: within two to five years, every organisation will be on a journey toward this new organisational form.

McKinsey’s 2025 State of AI research shows the same pattern from another direction. AI use is now widespread, but enterprise-level transformation remains immature. Many organisations are experimenting, piloting, and deploying tools, yet most have not embedded AI deeply enough into workflows and operating models to capture material value. In other words, the technology is spreading faster than organisational redesign.

That gap is the strategic danger.

The current adoption curve should not be mistaken for true preparedness. Using AI to write emails faster is not the same as becoming an AI-enabled organisation. Giving staff access to chatbots is not the same as redesigning decision rights, workflows, risk controls, knowledge management, customer journeys, operating rhythms, and performance systems around persistent agentic capability.

The future belongs to organisations that understand the difference.

Hard data already shows that AI is not a marginal productivity tool. Stanford’s 2025 AI Index reports that AI performance continues to improve across demanding benchmarks and that emerging studies confirm productivity gains, often narrowing the gap between lower-skilled and higher-skilled workers. Anthropic’s Economic Index, based on millions of Claude conversations, found that AI usage is concentrated heavily in software development and writing, but extends across roughly 36% of occupations for at least a quarter of their tasks. It also found that 57% of AI usage reflected augmentation, while 43% reflected automation.

That split is crucial. The lazy fear is “AI will replace everyone.” The more accurate reality is that AI will first recompose work. Some tasks will be automated. Some will be accelerated. Some will be de-skilled. Some will be up-skilled. Some will become obsolete. Some will become more human, more judgement-based, more relational, more strategic, and more supervisory.

The job does not disappear first. The task structure collapses first.

That is why the next serious workforce question is not “which jobs are safe?” It is “which parts of each role remain defensible when persistent agents can perform research, administration, reporting, communication, coding, monitoring, scheduling, document production, and operational follow-up?”

This is where Stage 2 becomes socially significant.

Persistent agents will not merely support work. They will expose the true design quality of work. Bureaucratic organisations built on repetition, handoffs, meetings, approvals, inbox traffic, status reporting, and document churn will find that much of their apparent complexity was never strategic complexity. It was (and is) coordination waste and so costly.

The uncomfortable truth is that many white-collar systems have confused activity with value. Persistent agentic systems will make that confusion harder to hide.

In the exponential age, the premium shifts from information access to judgement, orchestration, ethics, adaptability, evidence discipline, systems thinking, and creative recombination. People who can define problems, frame decisions, validate outputs, challenge assumptions, govern risk, and connect technology to real-world outcomes become more valuable. People whose work is primarily procedural, repetitive, approval-based, or template-driven become increasingly exposed.

That does not mean society should celebrate disruption blindly. It means society must prepare honestly. Education systems still built around single-loop delivery, memorisation, post "learning" assessment, mass delivery, chalk and talk, timetable fragmentation, age-batching, compliance tasks, and standardised progression are [not] preparing young people for a disappearing work structure. The exponential age requires something different: AI fluency, data literacy, agent management, ethical reasoning, triple loop, experiential, immersion and project-based learning, industry certification, entrepreneurial capability, and triple-loop learning, where learners do not merely ask “did I get the answer right?” but “what problem am I solving, why does it matter, what assumptions am I using, what system am I operating within, and how do I adapt?”

That is not futuristic educational theory. It is workforce survival.

By 2027–2028, the useful version of Stage 2 will be normal inside competent businesses. Inbox triage, meeting preparation, proposal drafting, CRM updates, customer follow-up, reporting, policy monitoring, risk scanning, document production, compliance preparation, research synthesis, and light operational execution will increasingly be handled by semi-persistent AI agents. These systems will not be perfect. But they will be good enough to fundamentally change expectations.

Once a business can produce in two hours what previously took two days, the market does not politely preserve the old rhythm.

By 2029–2031, the mature version of Stage 2 becomes more plausible: a digital chief-of-staff layer that remembers priorities, tracks projects, understands and maps stakeholder tone over time and situationally, manages workflows, monitors risks, drafts outputs, follows up, and operates within delegated authority. It will not be conscious. It will not be human. It will not have moral agency. But it will have operational continuity.

And operational continuity is enough to reshape society. It can be programmed now though.

This is why the phrase “artificial intelligence” is increasingly wrong. It directs attention to the wrong comparison. It invites people to ask whether the system is like a human mind. But the most important systems are not trying to become human minds. They are becoming synthetic cognition infrastructure.

A better definition would be -

Artificial intelligence is machine-enabled cognition that can perceive, interpret, generate, decide, remember, and act within digital or physical systems to achieve human-defined or system-defined objectives. Even that definition may need sharpening. For the Stage 2 world, the more useful phrase is probably not artificial intelligence at all. It is “synthetic operational intelligence.”

That phrase is more accurate because the real transformation is not artificiality. It is operationalisation. Intelligence is being embedded into work itself. Into documents. Into workflows. Into devices. Into vehicles. Into classrooms. Into hospitals. Into call centres. Into software development. Into government services. Into compliance. Into logistics. Into infrastructure. Into personal administration. Into every coordination layer previously dependent on human cognitive labour.

The old definition of AI makes people ask, “Can it think like us?”

The new definition forces a better question: “What can it now do that previously required organised human cognition?”

That is the question every board, school, agency, business owner, executive team, and worker must confront.

The societal implications are profound.

First, productivity will become more uneven. Organisations that integrate persistent agents into their operating model will compound advantage. Organisations that treat AI as a side-tool will become slower by comparison. The divide will not be between AI users and non-users. It will be between AI-operating organisations and AI-sprinkled organisations.

Second, trust becomes infrastructure. Persistent agents require memory, permissions, audit trails, identity controls, data governance, privacy boundaries, retrieval discipline, and human oversight. Without these, agents do not create intelligence. They create automated confusion at scale. Memory drift, stale context, hallucinated authority, poor access control, and opaque decision logic become operational risks.

Third, management changes. The manager of the near future will supervise hybrid teams: people, agents, automated workflows, external platforms, data sources, and machine-generated outputs. The core skill becomes orchestration. Leaders will need to decide what agents can do, what humans must approve, what evidence is required, where the audit trail sits, and when automation must stop.

Fourth, work identity changes. People have long been valued for knowing, remembering, drafting, coordinating, and reporting. Persistent agents attack those identity zones directly. This will create fear, resistance, grief, status anxiety, and institutional denial. Good leadership will not mock that fear. It will redesign roles around higher-value human contribution.

Fifth, national preparedness becomes a strategic issue. As deglobalisation continues, Countries that fail to modernise education, digital infrastructure, public-sector systems, procurement models, and workforce transition pathways will experience AI as external pressure rather than internal capability. They will buy tools but not build capacity. They will regulate symptoms but not develop competence. They will talk about innovation while preserving systems designed for the pre-agentic age.

That is the real risk.

The exponential age does not reward passive adoption. It rewards preparedness.

Preparedness means every organisation should now be asking seven hard questions.

What work do we perform that is actually repetitive cognitive labour?

What decisions require human judgement, and which merely require process execution? Where is our data too fragmented for agents to operate safely?

What permissions would an agent need, and what should it never be allowed to access? What workflows could be redesigned around human-agent collaboration?

What skills must our people develop to supervise, validate, and direct agents?

What operating model would we build if persistent intelligence were assumed rather than exceptional?

Rare Digital Innovation’s view is that Stage 2 is not a distant technology event. It is an operating-model event. It is the beginning of the shift from digital tools to digital colleagues, from software as system of record to software as system of action, from passive information systems to active intelligence systems.

The timeline is brutally short.

Stage 2 is perhaps 10% here now. It may be 50% useful by 2027–2028. It may be 80% mature by 2029–2031. Those percentages are not scientific certainties. They are strategic markers. They say: the window for preparation is now.

The winners will not be those who worship AI. They will be those who govern it, shape it, embed it, challenge it, and redesign work around it.

The losers will not necessarily be those who reject AI loudly. They will be those who adopt it shallowly or misunderstand it.

The next stage of artificial intelligence is not consciousness. It is continuity. It is memory plus action. It is persistent cognition connected to tools. It is synthetic operational intelligence becoming part of the fabric of work.

That is the real direction of travel.

And for any organisation still asking “how far away is this (and can we shuffle our paper about for a while longer to ignore it all)?”, the more useful answer is:

Close enough that delay is now a strategy. Just not a good one and never was.