Architecture of Action
in the Agentic
Era

The market is not breaking

The old unit of professional value is breaking.

In every technological wave, people feel as if the system has stopped working. Old roles lose clarity, familiar prices no longer explain effort, and new participants produce results that once required a large team. It becomes tempting to say: “the market has been destroyed.”

But markets rarely disappear completely. More often, work is broken down into tasks, some tasks are automated, some become cheaper, and the profession is assembled again around new scarcities.

OpenAI’s model of the jobs transition proposes four paths: higher automation potential, reorganization of the occupation, increased demand because costs fall, and relatively limited near-term change. In the United States, these categories include about 18%, 24%, 12%, and 46% of jobs respectively; the European extension estimates the groups at roughly 14%, 27%, 12%, and 47%. OpenAI emphasizes that this is not a forecast of mass job disappearance, but a map of different kinds of adaptation.¹²

Developers, accountants, lawyers, teachers, and financial analysts do not necessarily disappear. Their work is reorganized: AI takes on more standardized operations, while people remain central where judgment, responsibility, relationships, physical presence, understanding exceptions, and accepting consequences are required.

In the past, a large share of a specialist’s value was held in the direct production of an artifact. Now the first version is more accessible. But access to a draft is not the same as access to a high-quality result.

The vacuum cleaner did not abolish cleanliness. AI does not abolish engineering. It raises the standard.

Household technologies reduced the labor intensity of individual operations, but at the same time changed expectations for cleanliness, frequency of care, domestic comfort, and quality of life. Similarly, AI reduces the time required to produce a digital result, and the market starts to expect more: faster, safer, clearer, more personal, and better documented.

Not human versus AI

The right unit of competition is neither the person nor the model. It is an entire system of action.

The opposition between humans and artificial intelligence works well for headlines, but poorly for designing reality. People and models have different strengths. A system becomes stronger when functions are distributed deliberately.

In its first Economic Index, Anthropic found that real Claude usage was more often amplifying than replacing work: 57% of observed usage was augmentation and 43% was automation. At the same time, AI affected not whole professions, but different shares of tasks inside them.³

This proportion is not permanent: agentic interfaces increase the amount of delegation. But greater autonomy does not remove the person. It changes the point of human involvement: people increasingly do not type every step, but define the goal, grant access, set checkpoints, verify the result, and resolve exceptions.

OpenAI’s research on Codex shows a transition from consultation to delegated production. In the first half of 2026, the number of active Codex users grew more than fivefold; more than 10% of users managed three or more agents in parallel at least once a week, and the share of users who submitted at least one task with a complexity above eight hours of human work grew almost tenfold.⁴

This is an engineering position. Complex systems require distribution of functions, observability, routing, feedback, and a responsible center for decision-making.

SOT × SAF

First see the field. Then make the precise move.

SOT × SAF is not a mathematical equation in the strict sense. It is a discipline of thinking: if a critically important dimension is zero, the whole result can collapse to zero. A team can be strong but choose the wrong time; see an opportunity but lack the skills; build a perfect system that does not fit the user.

Space

Space answers the question: where does the action actually happen? Space includes infrastructure, channels, law, language, trust, culture, data, geopolitics, energy, platforms, and institutional rules.

Opportunities

Opportunities answers the question: what has become possible right now? An opportunity appears when the combination of price, speed, availability, maturity, and need changes.

According to Epoch AI as of February 2026, inference cost at a fixed quality level was falling by roughly 40 times per year; the total compute capacity of the AI chip fleet was growing by about 3.4 times per year; compute used to train frontier language models by about five times; and pre-training efficiency by about three times annually.⁵

Time

Time answers the question: when should we act? A technology can be right but too early. A product can be useful but already late. Speed can create an advantage, or it can make a team move quickly in the wrong direction.

Skills

Skills are not only personal knowledge. They are the full set of capabilities available to the system: a person, a team, an external specialist, a model, an agent, a tool, a library, data, capital, and infrastructure.

Alignment

Alignment answers the question: can the parts of the system work together? Even strong elements produce weak results if goals, incentives, definitions, and responsibility diverge.

Accenture describes agentic transformation through abundance, abstraction, and autonomy, and emphasizes that autonomy requires new mechanisms of trust, monitoring, and governance. 77% of surveyed executives expected agents to reinvent the creation of digital systems, and 78% expected digital ecosystems to have to be built for agents as well as for people.⁶

Fit

Fit answers the question: does the solution fit these people, in this place, at this moment? Technical strength is not the same as appropriateness. Sometimes the best architecture is the one the team can maintain. Sometimes the best product is minimal.

Focus, locus, and context

Focus increases resolution inside the chosen area, but reduces the field of view.

Focus is necessary. Without it, it is impossible to build a product, finish a study, or make a decision. But focus is not neutral: it always cuts off part of reality.

What lies outside our focus does not stop existing.

Outside the beam of attention remain weak signals, side effects, absent specialists, alternative solutions, and changes in the environment. That is why a professional must ask two questions: “What are we focused on?” and “What have we stopped seeing because of that?”

Focus determines precision. Leverage shows where a real lever exists. Context prevents a locally precise but systemically wrong decision.

Three circles

In a financial environment, it is especially dangerous to confuse ownership with control. A person can legally own an asset without controlling its liquidity, issuer, rules of circulation, infrastructure, or possibility of exit. Similarly, a company may think it owns a digital product while being fully dependent on a single API, app store, or cloud provider.

AI compresses execution, but it does not return lost time

Agentic systems create the illusion that time has become infinite. The model does not sleep. Several agents can work in parallel. A task that used to take a week can sometimes be done in hours.

But this is machine time. Human attention remains limited. The market window closes. A wrong hypothesis keeps burning money. A bad decision produced ten times faster does not become good.

The main production unit of the future is not a line of code or a single answer. It is the speed of moving through the cycle:

A team that generates faster can lose to a team that learns faster. The agentic economy increases the value of feedback: tests, metrics, observability, customer signals, review, and the ability to roll back.

History repeats not events, but patterns of transition

Gutenberg: not a new book, but a new knowledge economy

The printing press matters not because it made it faster to copy text by hand. It changed the cost of copying, the speed at which ideas spread, the number of people involved in producing knowledge, and the possibility of mass verification. Research links earlier diffusion of printing in European cities with later growth in the number of notable scientists and artists.⁷

The historical rhyme with AI is not that the technologies are identical, but that the marginal cost of producing and transmitting knowledge changes. When information becomes cheaper, power, education, markets, and criteria of authorship are reorganized.

Kodak: seeing the technology is not enough

The popular version says Kodak simply failed to notice digital photography. That is wrong. Kodak engineer Steve Sasson created a portable digital prototype in 1975; the company worked with digital technologies consistently. The problem was deeper: the transition required cannibalizing a profitable film economy and rebuilding incentives, channels, company identity, and the business model.⁸

It is not enough to see the future first. You must be able to rebuild the present.

The space pen: a beautiful myth and the right engineering lesson

The story that NASA spent millions on a pen while Soviet engineers used a pencil is a myth. The Fisher Space Pen was developed privately; NASA and the Soviet program used pencils at early stages, but graphite dust, broken tips, and flammable materials created risks, so both sides switched to special pens.⁹

The correct conclusion is stronger than the legend: a simple solution is good not because it is simple, but because it fits the real environment, constraints, and risks.

Radio amateurs, the airwaves, and modern creators

Early radio amateurs went on the air, broadcasting music, news, and conversations to small audiences. This could look like an unserious hobby, but those practices became prototypes for commercial radio and mass broadcasting.

Modern bloggers, independent developers, open-source maintainers, and AI-native teams perform a similar function: they create institutional forms before the market has learned how to name them properly.

The carrying yoke and running water

The carrying yoke was a rational technology within its system: it distributed load and improved water transport. Running water did not simply make the yoke more efficient. It moved the solution to a new architectural level.

In the same way, AI first accelerates an individual operation and then changes the whole system: processes, interfaces, organization of work, cost of coordination, and distribution of responsibility.

GitHub is not only code. It is a model of thinking.

Open source is simultaneously a technical, social, and epistemic technology. It lets distributed people find problems together, propose bounded changes, verify them, and preserve the history of decisions.

Issue requires naming the problem and providing context. Pull Request turns an opinion into a verifiable change. Review brings in someone else’s expertise. Merge means accepting responsibility for including the change in the main system.

This sequence can be applied to a product, organization, career, or personal life. Instead of breaking main, create a branch. Instead of an abstract intention, run a bounded experiment. Instead of self-persuasion, use review. Instead of an irreversible leap, make a verifiable merge.

But open source does not mean an absence of ownership, quality, or responsibility. It makes intellectual provenance visible: who proposed, who fixed, who checked, and which project a new branch grew from.

From intellectual property to intellectual provenance

Intellectual property is necessary as a legal mechanism for authorship, investment, and licensing. But it should not create the illusion that a result emerged in an intellectual vacuum.

Every modern product depends on language, mathematics, scientific institutions, electrical grids, semiconductors, standards, libraries, data, and the labor of people its creator will never meet.

That is why the question “who owns the result?” should be accompanied by another question: “which dependencies did it grow from?”

From assistance to delegation

An agentic system differs from a regular chat interface because it does not only explain, but also acts: reads files, calls tools, runs commands, changes artifacts, verifies the result, and continues the cycle.

OpenAI’s research on Codex shows that intensive users organize work around large, repeatable, and parallel processes. Inside OpenAI, by June 2026 Codex accounted for 99.8% of output tokens in work interactions across Codex and ChatGPT; among organizational users the figure was 63.3%, and among individual users 16.5%. The authors warn that OpenAI’s environment is not representative of an ordinary organization, but it shows a possible trajectory when adoption barriers are minimal.⁴

A model’s capabilities and an organization’s ability to extract value are not the same thing. Access, security, training, review processes, specifications, trust, and stop rules are needed.

AI 2027: a speed scenario

AI 2027 should be read not as prophecy, but as a concrete acceleration scenario. The authors state directly that they depict one of many possible paths and do not know the exact date of AGI; 2027 was their most likely single year at the time of publication, but their median estimates were later.¹⁰

The scenario’s main mechanism is the automation of AI R&D: systems help create the next generation of systems, potentially accelerating the research and development cycle.

AI 2040: a coordination scenario

AI 2040 Plan A is explicitly described as a recommendation, not as the most likely forecast. The authors propose international transparency in AI R&D and a managed slowing of the race. For Beyond IT, what matters is not the specific political prescription, but the method: technological scenarios must be tested through governance, power, trust, verifiability, and institutional coordination.¹¹

Three levels of agentic maturity

A constraint is not an advantage by itself. But it changes the direction of optimization.

Sanctions, export controls, closed APIs, and shortages of compute resources create real losses. They should not be romanticized. But strategic thinking requires seeing not only the prohibition, but also the new selective pressure.

When access to compute is limited, the value of efficiency, distillation, local inference, open weights, alternative hardware, and independent infrastructure rises. When an external service is unavailable, demand appears for a local ecosystem. When one platform closes access, the value of compatible protocols and portability increases.

The Stanford AI Index 2025 recorded a rapid narrowing of the gap between open and closed models, as well as China’s growing strength in publications, patents, and competitive open-weight systems. Open models increasingly compete not across all tasks at once, but through price, local deployment, privacy, adaptation, and industry specialization.¹²

Here it is important to distinguish open source from open weight. Public weights do not always mean open data, full training code, or a permissive license. Still, weight availability enables local audit, fine-tuning, adaptation, and reduced dependence on APIs.

The internet is changing its economic contract

Cloudflare and publishers point to a new conflict: an AI crawler can consume content without returning comparable user traffic to the source. In 2025, Cloudflare introduced default blocking of known AI crawlers for new domains and an experimental Pay Per Crawl model.¹³

This is not only a dispute about bots. It is a change in the economics of authorship, attribution, traffic, and payment. The old contract of the search internet sounded like this: “we index your material and return an audience to you.” The agentic internet can answer the user directly, so new mechanisms will be needed for permission, licensing, citation, and value distribution.

From T-shaped to M-shaped and W-shaped

A T-shaped specialist has one deep domain and a broad outlook. This model remains useful, but the agentic era requires several deep supports and the ability to switch scale.

M-shaped and W-shaped are used here as the authors’ framework, not as a single generally accepted HR standard.

Do not lose the junior inside yourself

A junior is important not because of a lack of knowledge, but because of the ability to ask questions, acknowledge uncertainty, try things, and quickly update their view of the world. Professionalism should not destroy this ability.

The curiosity of a junior. The depth of a professional. The responsibility of an architect. The coordination of an orchestrator.

The full stack of the future is not a person who personally does everything. It is a specialist who can understand several layers of the system well enough to distribute work correctly, see interface conflicts, and make a decision.

Refactoring knowledge

In programming, refactoring changes the internal structure of a system while preserving or improving useful behavior. In thinking, refactoring means updating models of the world when new data appears.

We do not definitively know how consciousness arises. We do not understand the brain as a whole. We cannot treat the absence of a known function as proof that no function exists. Scientific maturity is not constant doubt about everything, but a precise understanding of the boundaries of evidence.

“Junk DNA”

Only about 1–2% of the human genome encodes proteins. Non-coding DNA includes many functional elements, such as regulatory sequences, non-coding RNAs, centromeres, and telomeres, but it also includes regions that may be nonfunctional. Therefore both simplifications are wrong: “all non-coding DNA is junk” and “all of it is necessarily functional.” The scientific conclusion is that we must distinguish absence of protein coding, biochemical activity, and evolutionarily meaningful function.¹⁴

The appendix, tonsils, and medicine

Organs that were once often described as almost useless have immune and physiological functions. This does not mean surgery is never needed. It means the decision should account for a richer model of the system and specific medical indications.

Tooth regeneration

Research into stimulating tooth growth is indeed advancing. Japan’s TRG035 program, which targets the USAG-1 protein, has passed an early safety study, but the ability to reliably restore functional human teeth remains experimental, not routine medicine.¹⁵

On the shoulders of titans

No modern achievement starts from nothing. A developer uses mathematics, physics, electrical grids, semiconductors, the internet, operating systems, protocols, libraries, documentation, and other people’s mistakes.

Beneath human civilization stands the first and greatest titan: nature. We have learned to use its forces locally, but we do not control the system as a whole. We do not abolish gravity; we build an airplane within its laws. We do not abolish thermodynamics; we design computing systems with energy and heat in mind.

The Kardashev scale is useful as intellectual calibration: humanity has not yet reached a Type I civilization capable of using planetary-scale energy as an integrated system. Despite neural networks and spacecraft, we remain an early technological civilization.

We are less like masters of the universe than like ants that have read the diagram of their anthill for the first time and decided they understand the whole forest.

AI does not make us omniscient. It gives us new optics. A telescope lets us see farther, but it does not choose the direction. A microscope magnifies the object, but it does not replace the scientific method. A model expands the information field, but it does not free people from verification and responsibility.