Designing for agency

As Verity entered 3.7, the oldest and most extensively pierced of the baristas shoved a dirty chai in her direction, across the zinc counter.

"I ordered for you," the voice expecting to be called Eunice said. Verity had covered the headset with a beanie she hoped wouldn't suggest she was trying to look younger. She decided to keep it on. "Thanks. How'd you know what I'd want?"

"Your Starbucks rewards account," said Eunice, so-called, practicing what she said was facial recognition on the barista. A right geometry formed, the cursor having found his face, straight lines connecting, centered around the sinus region, to zero in on the nose tip, and then was gone. This had started on the street, on the way over, though Eunice claimed to have no idea how she was doing it.
– William Gibson, Agency (2020)

In the context of the argument made here, by (AI) agent¹, I'm assuming the following simplified definition:

An autonomous system that perceives, decides, and acts in a loop to achieve goals in its environment.

This should be considered within the broader domain of agency², here simplified as:

The capacity to set goals, choose actions, and exert causal influence to realise them—subject to context, capabilities, and constraints.

Which within the broader context of AI and more generally defined Computational Agents, is very well rendered in Poole and Mackworth's

"AI is about practical reasoning: reasoning in order to do something. A coupling of perception, reasoning, and acting comprises an agent.
An agent acts in an environment. An agent’s environment may well include other agents. An agent together with its environment is called a world."
–David L. Poole and Alan K. Mackworth, Artificial Intelligence: Foundations of Computational Agents,  2nd Edition

---

The Design of Delegation and the De-Individualisation of Agency

For decades, digital system design has been fundamentally preoccupied with user interfaces, addressing the core question of how humans interact with a given software application.

The emergence of agentic artificial intelligence represents a qualitative shift that renders interface mechanics secondary. This shift can be made apparent by looking at the evolution of software behaviour:

• Tools: require constant human actuation (e.g. a spreadsheet).
• Assistants: execute specific, tightly bounded commands (e.g. summarising a document).
• Agents: perceive context, define intermediate steps, and execute actions to achieve a high-level goal, adapting to feedback along the way.

An AI agent is technically defined by:

• its active perception-decision-action loop
• operating with goal-directed autonomy to sense its environment
• reason through ambiguous parameters
• plan tasks
• use software tools with minimal real-time human intervention.

As software graduates from merely answering questions to independently executing multi-step workflows, corporate strategy must pivot toward the design of algorithmic delegation.

The central challenge of the modern enterprise is no longer structural layout, but rather defining which AI agent may act, under what authority, within what specific bounds, and with what operational accountability.

To understand how software networks can legitimately wield this administrative power, organisations must draw on foundational insights from institutional sociology and organisation studies.

Institutional theory demonstrates that agency must be completely dissociated from human individuals. Far from being an inherent, biological trait of atomised actors, agency is a systemic capacity or quality that stems directly from the resources, rights, and obligations tied to institutionally engineered roles.
When an enterprise maps these institutionally engineered roles onto interconnected agent networks, it establishes a computational bureaucracy.

This structure represents the modern instantiation of Max Weber’s ideal administrative apparatus, designed to execute operations sine ira et studio (without passion or prejudice) to eliminate human bias, enforce strict standard operating procedures, and maximise objective consistency, speed, and calculability.

The decisive reason for the advance of bureaucratic organization has always been its purely technical superiority over any other form of organization. The fully developed bureaucratic mechanism compares with other organizations exactly as does the machine with the non-mechanical modes of production.

Precision, speed, unambiguity, knowledge of the files, continuity, discretion, unity, strict subordination, reduction of friction and of material and personal costs--these are raised to the optimum point in the strictly bureaucratic administration, and especially in its monocratic form. As compared with all collegiate, honorific, and avocational forms of administration, trained bureaucracy is superior on all these points. And as far as complicated tasks are concerned, paid bureaucratic work is not only more precise but, in the last analysis, it is often cheaper than even formally unremunerated honorific service.
-- Max Weber, Economy and Society (1921)

When Weber stated that a fully developed bureaucracy compares with other organisations “exactly as a machine compares with non-mechanical modes of production,” he was somehow predicting the technical evolution we are witnessing today with agents. The transition from human-managed corporate hierarchies to automated, agentic workflows represents the final conversion of social bureaucracy into compiled, deterministic code.

The challenges of modern AI system design—building alignment guardrails, mapping explicit scopes, ensuring audit trails, and designing escalation paths—are the exact same organisational problems that states and corporations spent the last two centuries solving via paper files and civil service rules.

---

The Digital Bureau
Mapping Weberian Administration to Multi-Agent Systems

The structural parallels between multi-agent system architectures and classical bureaucracy are direct, systematic, and codifiable. By translating Weberian administrative theory into programmatic terms, the system architecture maps corporate controls into explicit digital components:

Weberian Bureaucracy Component	Agentic System Parallel	Operational Manifestation
Demarcated Jurisdictions	Tool Permissions & API Gateways	Strict least-privilege execution scopes that restrict an agent's ability to read, write, or call external system services.
Exhaustive Standard Operating Procedures (SOPs)	Prompt Engineering, System Instructions & Workflows	Deterministic guardrails, immutable behavioral rules, and execution loops built into the agent's core model context.
Hierarchical Supervision	Guardian Agents & Enforced Human Checkpoints	Continuous algorithmic auditing of intermediate reasoning steps, paired with mandatory manual verification for high-impact actions.

This alignment leads directly to an Ontological Transformation of the Official. Max Weber characterised the bureaucratic official as an actor who exhibits impersonal loyalty to a position, views their office as a vocation governed by strict regulations, and maintains a radical separation between personal interests and administrative duties.

In a multi-agent network, the official is no longer a human worker, but the runtime execution block itself, the digital agent instantiates the ideal Weberian bureaucrat: it possesses no personal psychological biases, its sole vocation is the continuous execution of its coded instructions, and its separation of spheres is absolute, ensuring that tasks are handled with uniform objectivity.

However, an enduring law governs this architectural transition: capability cannot scale faster than control.

Building a high-growth, automated enterprise depends on a disciplined convergence of corporate strategy, data architecture, and runtime system oversight.

---

The Crisis of Authority in Algorithmic Governance

While a digital bureaucracy optimises operational throughput, its mechanical nature introduces a distinct set of systemic governance vulnerabilities, structurally defined as a crisis of authority:

The Dilettante Sovereign
A critical breakdown in administrative oversight occurring when human managers are tasked with supervising complex agentic workflows without understanding the underlying model logic.

Lacking technical visibility, these supervisors resort to superficial rubber-stamping, blindly approving automated outputs and turning human-in-the-loop controls into an empty compliance exercise.

The Cryptographic "Official Secret"
The acute explainability crisis that arises when multi-agent networks route reasoning paths across high-dimensional vector spaces, hidden neural weights, and multi-layered API calls.

Because the internal rationale is computationally opaque, the system creates a digital version of the bureaucrat's guarded secret, making post-hoc auditing incredibly difficult for external stakeholders.

The High-Frequency Capitalist Exception
Systemic friction caused when data anomalies, boundary mismatches, and API failures propagate across interconnected agent networks at mechanical latencies.

These errors compound far faster than the response bandwidth of human administrative structures, leading to rapid exception loops or silent data corruption.

This operational environment marks a shift from collegiate deliberation to monocratic execution. Traditional, slow corporate committee theatre and shared group-thinking are replaced by rapid, monocratic machine decisions. To maintain systemic legitimacy, these automated actions cannot rely on human intuition; they must be verified by explicit algorithmic certificates, automated model evaluations, and immutable alignment weights embedded directly within the execution stream.

Moving past isolated pilot models requires embedding governance directly into the execution environment.

---

The Structural Evolution

Just as a classical Weberian administration maintains a graded hierarchy of officials with varying levels of administrative competence, a computational bureaucracy evolves across distinct maturity plateaus.

To prevent the strategic blind spots of methodological individualism—the flawed paradigm that treats an AI agent as an isolated, willful actor with an inherent capacity to manipulate processes—enterprise capability development must be measured against a rigorous socio-technical maturity matrix.

Promotion between these structural tiers is strictly blocked unless the execution layer fulfils the corresponding Agent Experience (AX) and safety criteria:

Level 0: Decorative AI (Passive Assistance)
Pre-agentic manual execution where software acts as a passive administrative tool, offering basic transactional utility (e.g., drafting job descriptions or parsing raw log files) while human operators retain total operational control. No explicit autonomy model exists.

Level 1: Assistive AX (Localised Suggestion)
Semi-autonomy where the system processes localised data to propose discrete, single-step options (such as cash-flow forecasting or low-stock alerts), leaving definitive validation to the human supervisor. System state tracking and internal reasoning remains opaque.

Level 2: Controlled Delegation (Bounded Execution)
Contextual autonomy, marking the functional baseline of true agentic software. The agent independently executes tightly bounded, end-to-end repeatable tasks (e.g., accounts payable three-way matching) within closed corporate environments under explicit, hardcoded recovery constraints and mandatory verification checkpoints.

Level 3: Governable Autonomy (Adaptive Orchestration)
Domain autonomy where multi-agent networks coordinate complex, multi-step workflows with deep structural dependencies. The system dynamically manages standard process variations, routing true exceptions to human supervisors. Operations run within an active framework of programmatic Policy-as-Code, guardian agent oversight, real-time contestability, and clear escalation ownership.

Level 4: Protocol-Native Organisation (Strategic Self-Evolution)
Full strategic autonomy where the organisation’s core services, compliance boundaries, and assets are fully represented through machine-readable semantics. Operations are steered by a self-governing, self-improving network of collaborating agents that safely negotiate, trade, and reallocate corporate assets across independent ecosystems using open communication protocols.

Viewed from an institutional perspective, each step up this matrix represents a progressive, programmatic reallocation of organisational rights and obligations to machine actors occupying engineered corporate positions.

Unmanaged autonomy at advanced levels creates a profound operational crisis; without structural containment, autonomous systems generate excessive exceptions and allow errors to propagate silently across interconnected corporate applications. This creates an administrative bottleneck, forcing continuous human intervention and diluting the efficiency and speed advantages that motivated the system’s deployment.

Architectural Integration Tracks and the Systemic Security Perimeter

To safely deploy a computational bureaucracy, an organisation must align its technical infrastructure with its broader tolerance for operational risk. Multi-agent integration follows three primary corporate tracks:

Agents overlay
Intelligent capabilities are added as external enhancements to legacy software systems and pre-existing workflows, delivering fast productivity gains without requiring a foundational overhaul of underlying systems.

Agents As-a-Service
Pre-packaged agentic functionalities are utilized directly out-of-the-box from cloud software providers, allowing rapid internal scaling while binding compliance monitoring to vendor boundaries.

Agents by design
Core workflows and systems are re-architected from scratch as native networks of collaborating autonomous agents, building security boundaries and behavioural rules directly into the digital fabric of the business.

Transitioning from an incremental overlay to a native by design network marks a critical shift from localised model risk to comprehensive system risk.

In a high-agency multi-agent environment, the traditional boundary separating information security from model alignment completely collapses. A safety failure and a cybersecurity breach become the exact same operational event.

For example, an adversarial prompt injection attack is simultaneously a security exploit (unauthorised code execution) and a safety failure where the agent breaks out of its institutionally engineered role boundaries.
Similarly, data retrieval failures or hallucinated parameters that trigger erroneous downstream API transactions violate both security controls and operational integrity.

Consequently, the architecture must operate under a strict untrusted by default design pattern. No agent output can be accepted blindly; every machine-generated decision that initiates a financial transaction, alters a core database, or drives an unmoderated customer interaction must be programmatically verified by an independent validation layer prior to execution.

---

Policy-as-Code

The principal obstacle preventing enterprises from scaling autonomous operations is the implementation gap—the strategic failure to convert high-level corporate governance guidelines (such as fairness, transparency, and data privacy) into enforceable, deterministic technical controls at runtime. A computational bureaucracy cannot function on soft guidelines; it requires its rules and techniques of control to be hardcoded into the execution environment.

This enforcement is achieved by translating corporate compliance parameters into machine-readable and machine-enforceable rules via policy-as-code. Delegation-of-authority limits, budgetary boundaries, and risk appetites are compiled into real-time validation filters that monitor the agent's input and output channels. This programmatic approach provides continuous enforcement, injects real-time guardrails into multi-agent execution graphs, and produces the forensic, immutable audit trails required to satisfy rigorous regulatory compliance frameworks.

Furthermore, because autonomous networks operate at latencies that outpace human observation, the architecture must introduce specialised supervisory systems known as guardian agents. These independent systems are engineered specifically to monitor, validate, and manage operational agents at scale. Guardian agents continuously evaluate intermediate reasoning steps, check data access lineages, and audit generated outputs against corporate compliance standards. If an operational agent experiences a logic breakdown, attempts an unauthorised API call, or encounters a high-impact edge case, the guardian agent triggers an immediate intervention or routes the process to a human supervisor.

---

Socio-Technical Infrastructure and the Agent Mesh

In classical organisation studies, collective action relies heavily on institutional scripts—observable, recurrent patterns of activity and interaction that structure context-relevant behaviour within a particular setting. For a multi-agent system, this interactional order requires a highly unified, reliable socio-technical infrastructure; if data pipelines are fragmented or lineage is unverified, the underlying autonomous agents inherit those operational weaknesses, breaking the behavioural scripts of the computational bureaucracy.

To prevent these systemic blind spots, enterprises must build convergence architectures that unify live operational data streams and deep analytical records onto a single cloud-native data platform. By linking streaming, event-driven technologies with a vectorised data fabric, agents receive low-latency, context-rich access to real-time enterprise information. This data foundation acts as the organisation's objective knowledge discourse, enforcing data lineage as code and giving agents a deterministic way to verify high-trust data sources while providing auditors with transparent provenance to combat bias and hallucination.

Concurrently, the interaction framework of the enterprise must evolve away from rigid, point-to-point API connections, which generate fragile code dependencies and system bottlenecks. To scale autonomy sustainably, organisations must adopt standardised communication frameworks that codify modern organisational scripts:

• Model Context Protocol (MCP) & Agent-to-Agent (A2A) Frameworks: These open protocols provide uniform methods for agents to safely discover tools, interoperate, and exchange context across distinct application environments without requiring hardcoded integrations.
• The "Agent-Mesh" Orchestration Layer: The organisation treats its fleet of digital agents like a cloud-native microservices mesh, establishing a formalised interaction order. This mesh provides centralised security discovery, token exchange, and detailed telemetry across the entire operational network.

This infrastructure operationalises a comprehensive three-tiered governance architecture:

1. Enterprise Governance clarifies ultimate accountability, defining who owns systemic risk and holds the authority to approve or abruptly deactivate live agent meshes.
2. Lifecycle Governance institutes strict, evidence-based code gates and adversarial red-teaming protocols before agents can enter production.
3. Ecosystem Governance secures the perimeter against external model providers and third-party data APIs, ensuring that chained agent dependencies do not introduce vulnerabilities into the core organisation.

---

Redefining Work and the Return-on-Autonomy

The macro-economic impact of the computational bureaucracy marks a profound shift in corporate capital allocation. Agentic AI represents a new economic resource termed cognitive capital. By enabling capital assets to directly augment or substitute for cognitive tasks like never before, agentic AI is changing the rules of enterprise competition and shifting the economic value equation away from labor inputs and heavily toward capital ownership.

Agentic AI marks a shift in economic history—By enabling capital to augment or substitute labor like never before, AI agents are changing the rules of competition and redefining the balance of economic value creation.

(...)

Agentic AI is no less than a new type of capital that companies can leverage to generate value, the same way traditional capital (e.g., tools, machinery, know-how) has been used since the early days of capitalism. Sometimes referred to as “cognitive capital”, agentic AI allows businesses to re-think how they balance their factors of production, i.e., labor, natural resources, tangible and intangible assets, in their economic equation. Agentic AI is shifting the value equation further toward capital (and those who own and manage it) and away from labor.

(...)

Winners will place agentic AI where it unlocks 10x value, not 10% savings— The starting point is not tech or pilots: It’s building an enterprise portfolio of agentic AI opportunities.
– Six key insights for C-suite executives to maximize the ROI of agentic AI (2025), Accenture³

Strategic leaders do not deploy multi-agent networks for minor, incremental 10% cost savings; instead, they target 10x process reinvention by completely re-architecting end-to-end corporate workflows. This transition underpins the rapid expansion of the technology consulting sector, as enterprises rush to build comprehensive portfolios of agentic opportunities.

This shift does not eliminate the human workforce, rather, it redefines the corporate civil service through the lens of modular individuals. Human professionals step out from generalised positions into highly specialised, institutionally engineered roles focused on system architecture, risk absorption, and strategic orchestration:

Agentic Process Architect
Responsible for mapping the enterprise layout, identifying high-value agentic opportunities, and designing multi-agent workflows that align to corporate risk thresholds.

Autonomy Auditor & Continuous-Learning Steward
Tasked with forensic decision auditing, policy-as-code alignment reviews, adversarial red-teaming, and managing reinforcement learning data pipelines.

Human Escalation Officer / Functional Agentic Supervisor
Sits on the operational front lines, acting as the definitive administrative authority to handle complex exceptions, review high-risk outputs, and execute manual system overrides when safety parameters are breached.

To maintain strict institutional control over this hybrid environment, organisations must deploy elaborate measurement and documentation frameworks to track action outcomes. This is achieved via a Return-on-Autonomy (RoA) scorecard, which continuously monitors process-level and portfolio-level performance indicators:

Control Domain	Process & Portfolio Key Performance Indicators
Cost Takeout	• % reduction in total operating costs from baseline • % reduction in unit transaction expenses
Operational Speed	• Average weeks required to deploy new capabilities • Median process cycle times (hours/days compressed)
Productivity	• Aggregate output index per human FTE • End-to-end task autonomy rate (excluding overrides)
Quality & Safety	• First-pass yield (% tasks completed without rework) • Year-over-year reduction in regulatory findings
System Trust	• Agent reliability (% workflows completed without escalation) • In-workflow user and stakeholder satisfaction scores

• Cost takeout: Tracks percentage optimizations in overall corporate operating budgets and parallel drops in unit transaction expenses.
• Operational speed: Measures average deployment latencies and workflow cycle compression, charting how rapidly the system adapts to market shifts.
• Productivity: Evaluates total output per human employee alongside the absolute task autonomy rate (the percentage of corporate activities finalized end-to-end without manual confirmation clicks).
• Quality & Safety: Measures first-pass yield (the ratio of automated workflows executed with zero required manual rework) and tracks annual drops in compliance defects or regulatory audit findings.
• System trust: Continuously instruments the agent reliability rate (the proportion of autonomous workflows successfully completed without exception routing) alongside qualitative stakeholder satisfaction scores.

---

The ultimate competitive advantage in the agentic era does not belong to the institution that deploys the highest volume of isolated AI models, but to the enterprise that maintains the most rigorous control over how its autonomous systems interact, access data, execute tools, and obey corporate regulations.

True agentic scaling succeeds only when organisations abandon the myth of standalone model agency and recognise that agency is an engineered corporate capacity nested within structural perimeters.

By designing multi-agent orchestration as a structured, socio-technical operating system—grounded in policy-as-code, guardian agent oversight, and explicit, metric-driven techniques of control—organisations can safely build digital bureaucracies that are highly autonomous, structurally compliant, and completely resilient against systemic failure.

✺✺✺

---

Further reading

Artificial Intelligence: A Modern Approach, 4th US ed. (pdf)

Intelligent Agents: Theory and Practice (pdf)

Agents Situated in Environments (Artificial Intelligence 2E foundations of computational agents)

A Survey of LLM-Driven AI Agent Communication: Protocols, Security Risks, and Defense Countermeasures (pdf)

Artificial Intelligence: Foundations of Computational Agents,  2nd Edition

Agentic enterprise 2028 A blueprint for cost savings, job creation, and faster growth through agentic AI - Deloitte

---