The agentic extension of the thesis

Consider a limit case. If Distribuidora Norte, on its second pass, the one about delivery, handed route assignment to an agent without auditable rules, a failure would no longer be a salesperson ignoring an alert, a recoverable error. It would be a dispatch executed thousands of times with no witness. That shift of role is what this chapter is about.

None of the eight currents in Authors and currents ever saw an AI agent: Tavistock is from 1951, Latour from 2005 and Rogers’s diffusion from the sixties. The book holds that the thesis does not age with the technology, but until now it asserted this without working out what shifts when the technical system stops being an artifact that is used and becomes an actor that decides and executes on its own. This chapter builds that extension. It does not correct the eight currents but rereads them under agency. What it adds, which is nearly everything, are constructs of the paradigm, not attributable to any author. Where recent academic literature arrives at the same place on its own, it is flagged as validating convergence, not as the source this is derived from.

From the artifact that is used to the actor that decides

The bridge thesis was built on an implicit image of technology: an artifact that a person uses, operates or consults. A billing system, an ERP and even a predictive model do not decide: they record, sort, suggest. The decision, and with it the responsibility, always stays on the human side. The predictive model flags a customer as at risk, but it is the salesperson who decides to call. That frontier, where the human is the last filter of every case, is the silent assumption on which the whole method rests.

The agentic layer breaks that assumption. It is the rung of the adoption ladder at which software stops suggesting and starts to decide and execute end to end: it approves a refund, routes a complaint, prioritizes a case file or triggers an action without anyone looking at each case. The leap is not in the model’s power but in its role. The technical system stops being something that is used and becomes something that acts.

Here a gap appears that none of the eight currents fully fills. Actor-network theory already treated non-humans as actors: a breakwater, a speed bump or a key act in the network, enroll others, resist.¹ But Latour’s non-humans act without deciding. They do what their material inscription fixed for them and do not choose among courses of action according to a criterion they weigh in the moment. The theory has a vocabulary for the non-human that does, but not for the non-human that decides.

That gap is filled by what the paradigm calls the deciding actor. The AI agent is an actor in the actor-network sense, because it takes part in the network, enrolls and produces effects, but it belongs to a kind that theory did not typify: it decides among alternatives and executes its decision without a human filtering it case by case. It has no interests or intention, because it is not a subject, but it does exercise a delegated discretion: it operates within a margin of choice a human entrusted to it. The sociotechnical problem stops being “do people adopt this artifact?” and becomes “who answers for what this actor decides, and by what rules does it decide?”. This is a conceptual decision, not a finding. Extending the vocabulary of the actor preserves the symmetry that makes the theory powerful, with humans and non-humans in the same network, and adds only what is missing: the margin of decision.

The thesis: it does not break, it shifts

Faced with the agentic layer, the temptation is to declare the method obsolete. If the system decides on its own, what are the human reading, control and joint optimization for? This extension answers the opposite, and that answer is its thesis: what the paradigm calls the displacement of control. Neither Tavistock’s joint optimization nor Latour’s translation breaks when the actor decides; they shift one level upward. Human control stops being exercised over each decision and comes to be exercised over the rules by which the agent decides: what it resolves on its own, when it must escalate to a person, what it leaves on record and how it can be revoked. It is the move from governing the decision to governing the governance of the decision. The primacy of the human does not disappear; it changes its object.

The thesis was already latent, scattered across three places in the method. It appears in link 4 (“human control moves one rung up, from the decision to the governance of the decision”), in the concept judgment in the design, not in every transaction and in the potential dimension of IMIA. All three say the same thing from different angles; naming it the displacement of control makes it the axis of the extension.

It remains to specify what is preserved. Joint optimization held that the social and the technical cannot be improved separately. Under agency, the technical subsystem decides within the social one, and the interdependence does not loosen: it intensifies. An agent poorly coupled to the human fabric no longer produces a tool nobody uses, the classic and recoverable failure, but decisions actually taken that have to be undone. Translation, in turn, now includes translating human judgment into rules for the agent to execute. That act is itself sociotechnical, because it decides which judgment is delegated and which is not.

Each current, reread under agency

Actor-network. Translation in flesh and blood held that it is a person, not a document, who aligns the actors of an organization. The agentic layer adds an actor, the deciding one, and with it a twist: the agent is not only enrolled, it also enrolls. When it prioritizes case files, officials reorganize their work around that order; when it approves refunds, the fraud team adjusts its controls to what the agent lets through. That agentic enrollment is translation read in reverse. The organization does not only translate the world into the agent by giving it rules; once in operation, the agent translates the organization, because it sets the pace, the format and the distribution of the human work that remains around it. That is why the sociotechnical reading does not end when the agent goes into production: the agent changes the organization that had been read.

Human-centered AI. Shneiderman holds that high automation and high human control are not a trade-off.² The agentic layer seems to contradict him, since if the agent decides on its own, control should drop by definition. It does not, once it is clear where control is exercised. “High human control” stops meaning that a person reviews each decision, which at scale would be impossible or would reduce that review to a rubber stamp. It comes to mean that a person designs, audits and can revoke the rules by which the agent decides. Shneiderman’s quadrant holds; what moves is the point where control is applied, from the case to the rule.

Dignum and O’Neil come in with more force, not less. Accountability mattered when the model suggested, and it becomes structural when the agent decides, because it no longer falls on the person who pressed the button: there was no such person.³ The harm at scale that O’Neil documents also changes in nature. An opaque model that suggests badly produces suggestions a human can still discard; an agent that decides badly produces decisions already executed.⁴

Diffusion. The adoption ladder already places the agentic layer on its top rung; the rereading adds a nuance about two of Rogers’s perceived attributes.⁵ Observability is inverted: the better an agent works, the less it is seen, because it decides well, silently and without asking for anything. The attribute Rogers counted as a driver of adoption thus becomes a governance problem, since what is not seen is not audited. Trialability, the chance to try things out gradually, also gets harder when what is being tried executes: an agentic pilot is not harmless the way one that suggests is, because it already acts. The descriptive, non-normative use of Rogers that the method maintains becomes more demanding under agency. To spread an ungoverned agent quickly is to spread an actor that decides badly at the speed of the S-curve.

Tacit knowledge. Here the extension touches its most delicate point. The displacement of control assumes translating human judgment into rules, but Polanyi’s paradox, “we know more than we can tell”, indicates that the core of expert judgment is tacit and cannot be fully codified.⁶ If so, not all judgment can be turned into rules: what cannot be codified is what makes the expert an expert. From this comes what the paradigm calls the delegation frontier: the limit, shifting and always partial, between the judgment that can be carried over into auditable rules (thresholds, explicit criteria, typified cases) and the judgment that cannot, because it is tacit, contextual or loaded with conflicting values that no rule resolves without deciding for people. The frontier is sociotechnical, not technical. Defining what falls on each side is the human reading of the agentic layer, and designing an agent is, above all, deciding well where to draw it and obliging the agent to escalate to a person everything that falls on the tacit side.

Weick adds the last piece. If the organization enacts the categories by which it measures the world, an agent that decides on those categories does not merely reproduce the blind spots of constructed meaning: it executes them and, by deciding, reinforces them, because its decisions become tomorrow’s data.⁷

Governance rises with autonomy

From the above follows what the paradigm calls the proportionality rule: the governance a system requires is proportional to the margin of decision delegated to it. Automating a fixed-rule task, such as a report, an upload or a calculation, calls for little governance, because the system does not decide: it executes what is prescribed. An actor that decides among alternatives demands high governance, with usage policies, risk management, auditable rules and paths for escalation and revocation. Autonomy without proportional governance is not an opportunity but delegation in the dark.

This does not redefine IMIA; it grounds it. Its potential dimension already distinguishes fixed-rule automation from agentic automation and asserts that the agentic component raises the bar for the governance dimension. The proportionality rule is the general principle of which that distinction is one case. That is why the warning that “high agentic potential over low governance is not opportunity, it is risk” does not stand alone. And that is why IMIA’s governance gate, which prevents reaching the integration level with governance below the floor, is the strict form of proportionality: the higher up the ladder, the higher the governance floor for the leap to count as maturity and not as exposure.

Three risks of its own

The agentic layer does not gradually worsen the known risks. It introduces three that do not exist while the human filters each case.

Execution at scale without a witness. The human in the loop did not only filter: they saw, because each decision passed before someone’s eyes. When the agent decides on its own, that gaze disappears from the transaction and no one witnesses the individual decision. The risk is not speed, which already had a name (automating the error faster), but the absence of a witness. Governance has to restore, through logging, sampling and auditing, the witness the flow lost.

The error that repeats itself. A human who makes a mistake tends to notice and correct it. An agent that decides badly because of a poorly drawn rule repeats the same error in every identical case, without fatigue or doubt, until someone intervenes on the rule. That is why auditing the agentic layer watches the rule, not the case.

The dilution of accountability. When no one decided each case, the risk is that the question of who answers gets the reply “no one”: not the agent, which is not a subject, nor whoever designed it, nor whoever operated it. The antidote is the other face of the displacement of control. If control moves from the decision to the rule, accountability moves with it, and whoever designed, audited and maintains the rule is the one who answers. That is why those rules must be explicit, auditable and attributable to a person. Autonomy does not dilute human responsibility: it relocates it, and forces it to be made explicit precisely where it used to be implicit in the act of deciding.

The field converges

The constructs above come from within the method, from rereading the eight currents, and depend on no external source. Still, it is worth noting that, while this extension was being written, the academic literature reached the same reading on its own: agentic systems are sociotechnical systems. That convergence is not the source of the argument but its strongest validation. When works that do not know of one another describe the same object with the same apparatus, the object is real and not a rhetorical preference.

The first framework to treat agentic AI explicitly as a sociotechnical system closes its thesis almost in the words of this chapter: an agent’s behavior is co-produced by “algorithms, data, organizational practices, regulatory frameworks and social norms”.⁸ It is the deciding actor read as part of a network, and agentic enrollment, put in other people’s words. A second work formally updates the joint optimization of the Tavistock tradition for the age of AI and formulates it across four levels: individual, organizational, ecosystem and societal.⁹ That gives the thesis its version at scale, because an agent that decides badly does not stay within the organization that deployed it: it scales to the ecosystem.

The finding that most supports the thesis is empirical. Evaluated over twenty workflows, state-of-the-art manager agents still fail to jointly optimize goals, constraints and runtime.¹⁰ Put in the bridge’s vocabulary, joint optimization, coordinating the social with the technical, is what agents still do not do well. The sociotechnical reading is not a task agentic AI is about to absorb but the residue that resists automation.

Context as an asset

Between frontier agent engineering and the method there is an unexpected bridge, and naming it turns the artifacts of link 1 into infrastructure for the agentic layer. The emerging craft of building useful agents does not revolve around the model but around the context assembled for it: what it knows about the organization, in what vocabulary and with what memory of what has already happened. This is what has come to be called context engineering.¹¹ Read through the method, it is not a foreign technical novelty but what the bridge already produces. Institutional memory (recurring questions with consistent answers) and the project’s bilingual glossary (business and technical terms, defined once) are the context an agent needs in order to decide well.

From this follows a diagnostic proposition: an organization that cannot explain its process to a person cannot package it as context for an agent either. Organizational illegibility, once paid for in zero adoption, is now also paid for in the inability to delegate, because there is no context to give the agent. The usual order is inverted. One does not start with the agent and then “give it context”; one first builds the organization’s legibility, which is the bridge’s work, and that legibility is the context. That is why the book treats context as an organizational asset: the bridge does not only set the agent’s guardrails, it also manufactures the material the agent thinks with.

Why the thesis does not age

The book had been asserting that the bridge thesis does not age with the technology; this extension turns that into an argument. The agentic layer does not leave the sociotechnical reading without work: it changes its object and raises its stakes. Joint optimization and translation shift upward. Shneiderman’s control changes its point of application. Polanyi’s tacit judgment does not become codifiable; it becomes the frontier one has to know how to draw. Dignum’s accountability does not dissolve: it is relocated and has to be made explicit. Each rung of the ladder makes skipping the human reading more costly, and the agentic rung is the most costly of all, because at the top a process no one understood is not executed more slowly: it is executed on its own, at scale, in series and without a witness. Far from making the bridge obsolete, AI with agency makes it indispensable.

See also: Authors and currents · The bridge thesis · Concepts of our own · IMIA, the maturity instrument · Bibliography