Circo Studio · Lab Essays

ARL Manifesto · v1.1

Circo Studio Idea Lab

Open manifesto · v1.1 AI Responsibility Levels

A manifesto from the Lab

AI Responsibility Levels

A governance framework for collaboration between humans and artificial intelligence. Five levels. Three pillars. One conviction — autonomy is delegated, never assumed.

Preamble

We are uncovering better ways of working with artificial intelligence by doing it and helping others do it.

Through this work, we have come to value certain things over others. What follows is not a methodology nor a standard — it is a declaration of principles, written for those who design, implement, audit, or govern artificial intelligence systems within real organizations.

This document articulates three complementary frameworks developed in Circo Studio's Idea Lab, drawing from work in cloud architecture, enterprise governance, and digital transformation in Latin America's energy sector.

A note on form

This document takes its structure from the Manifesto for Agile Software Development (2001) — preamble, values stated as "X over Y" pairs, principles. The choice is deliberate and requires three honest asterisks before reading what follows.

I Difference of standing

The Agile Manifesto was descriptive: its seventeen signatories articulated what they were already doing, with years of consolidated practice behind them. This document also emerges from practice — we already govern AI implementations with this logic — but without that precedent's scale or consolidation. Its claim is more modest: to be a map, not a perfect guide. It declares routes we recognize, not closed paths.

II Difference of scale

Agile concerned itself with teams doing work. This document concerns itself with organizations governing systems. The form supports the borrowing, but the levels of concern are not symmetrical — what in Agile is an engineering practice, in ARL is a governance decision with regulatory, contractual, and reputational consequences.

III Ambiguity of the era

For part of the audience, Agile today evokes both the original spirit and its later bureaucratization — industrialized frameworks, certifications, agile theater. We take that load on knowingly. The alternative — inventing a new form to avoid contamination — would be worse: the reader would recognize the borrowing anyway, without our having declared it.

We value

We value —

Traceable responsibility over assumed autonomy

Human–AI symbiosis over pure automation

Cognitive efficiency proportional to value over maximum technical capability

Gradual and verifiable adoption over fast and ambitious deployment

That is — while there is value in the items on the right, we value the items on the left more.

Note — the counter-values are not caricatures. Pure automation, maximum technical capability, and fast ambitious deployment have value in certain contexts; we choose the left side for enterprise governance with real stakes.

Autonomy is delegated; never assumed.

The ten principles

Behind these values are ten principles that translate them into operational decisions. They are not commandments — they are convictions we use to discuss real projects with real clients.

Responsibility is never delegated all at once. It is graded by maturity, criticality, and verified trust.

Every AI action in a production system must be attributable, auditable, and reversible in proportion to its level.

The human who delegates remains responsible. AI is not a legal or ethical shield.

Before asking what AI can do, we ask what it should do and under what controls.

The level of autonomy is determined by the process, not the technology. The same model can operate at L1 for one flow and L4 for another.

Moving up a level requires evidence — telemetry, traceability, rollback mechanisms, and documented ethical review.

Human-AI collaboration is symbiotic when each side contributes what the other cannot, not when one replaces the other.

Business context precedes technical architecture. Governance is not designed in the abstract.

Every AI interaction consumes tokens, energy, and compute. Consumption should be proportional to the value generated, not to the urge to show off capability.

AI without governance is not innovation — it is deferred technical and reputational debt.

Circo Studio · Lab Essays

ARL Manifesto · Appendices

Circo Studio Idea Lab

ARL Manifesto · Sheet 2 Operational appendices

How it gets used.

What follows translates the manifesto's convictions into operational instruments — the five levels, the three frameworks that articulate them, and the terms teams use to discuss them.

Five levels, one shared language

ARL structures the human-AI relationship in five graduated levels, analogous to the Azure RBAC model. What RBAC does with permissions over resources, ARL does with decisions and cognitive actions — it defines what AI can do, in what scope, with what audit.

AI Responsibility Levels

Observation · consultation Human decides and executes. AI only suggests.

Exploration, early adoption, one-off prompts.

Supervised assistance Human review of every output before action.

Drafts, initial analysis, copy, assisted code.

Bounded collaboration Execution within guardrails. Sampling audit.

Agents with policies, Claude Code in scoped repos.

Delegated autonomy Human intervention only on exceptions.

End-to-end processes with telemetry and rollback.

Full autonomy Shared accountability. Continuous systemic controls.

Critical processes with multi-layer governance.

Criteria for moving up a level

Documented operational history at the current level.
Available and reviewed telemetry.
Tested rollback mechanism.
Ethics and compliance review.
Formal approval from the process owner.

The three frameworks

The manifesto articulates three complementary frameworks. Each answers a different question. None is sufficient in isolation.

Framework I

ARL · AI Responsibility Levels

Structures the human-AI relationship in five graduated levels, analogous to the Azure RBAC model. What RBAC does with permissions over resources, ARL does with decisions and cognitive actions. Answers the question of scope: what AI can do, in what scope, with what audit.

Framework II

Digital Symbiosis

Philosophical-operational framework of human-AI collaboration, structured in four pillars. Answers the question of quality: how that collaboration should operate within the autonomy contract.

Complementarity — each side contributes what's distinctive.
Transparency — the human knows what the AI is doing and why.
Traceability — every assisted decision remains reconstructible.
Mutual learning — the interaction improves outputs and human processes.

Avoids the two cartoonish modes of adoption — uncritical replacement, where things get automated without understanding what is being delegated, and decorative use, where AI shows up in presentations but not in the actual workflow.

Framework III

Cognitive Footprint

Principle of efficiency and environmental responsibility. Every interaction with a language model consumes tokens, energy, and compute. Those resources have economic, ecological, and cognitive cost. Cognitive footprint is measured and optimized like any other infrastructure resource.

Answers the question of cost: the chosen ARL level must be proportional to the value generated. A process that admits L2 is not run at L4 just because the model allows it.

The three frameworks operate as layers. ARL defines the autonomy contract. Digital Symbiosis defines the quality of the collaboration within that contract. Cognitive Footprint defines the admissible cost of that collaboration. The three, together, form a complete governance.

Operational use

In practice with clients, ARL works as shared language between business areas, IT, and compliance. It allows discussing which processes are candidates for which level, what controls are required to scale, and what evidence needs to be produced.

It is not an organizational maturity model — the same organization may operate at L1 in critical processes and L4 in bounded ones, simultaneously. The level belongs to the process, not the company.

Nor is it an external normative framework. It integrates with existing standards — ISO/IEC 42001, NIST AI RMF, EU AI Act — providing the operational language those frameworks leave to each organization's interpretation.

Terms

Accountability: Traceable responsibility for an action or decision. In ARL, accountability is never the AI's — it always falls on a human or an identifiable organizational structure.
Guardrail: Technical or policy restriction that limits the range of possible actions of an AI system within a defined scope. Prerequisite for L3 onward.
Cognitive footprint: Total cost — in tokens, energy, latency, and human attention — associated with an AI interaction. Central metric of the framework's third pillar.
ARL: AI Responsibility Levels. Five-level model that grades the autonomy delegated to AI systems according to criticality, maturity, and operational evidence.
Rollback: Documented and tested mechanism to reverse an action executed by an AI. Hard requirement for levels L3 through L5.
Scope: Bounded domain — temporal, functional, or data — within which an AI system operates at a given level. Outside the scope, the level does not apply.
Digital symbiosis: Mode of human-AI collaboration based on complementarity, transparency, traceability, and mutual learning. Second pillar of the framework.
Telemetry: Continuous record of inputs, outputs, latencies, errors, and decisions of an AI system in operation. Enabling condition for L3 onward.
Traceability: Property of an action or decision that allows reconstructing, post-hoc, which agent executed it, with what input, under what policy, and with what result.

Related frameworks and sources

Beck, K. et al. Manifesto for Agile Software Development. agilemanifesto.org, 2001. Structural inspiration for this document — see A note on form.
ISO/IEC 42001:2023 — Information technology — Artificial intelligence — Management system. International standard for AI management systems.
NIST AI Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, 2023.
Regulation (EU) 2024/1689 — Artificial Intelligence Act. European AI regulation, risk-based classification.
Microsoft Azure — Role-Based Access Control (RBAC). Reference model for the graduated structure of ARL.
Anthropic — Responsible Scaling Policy. Reference framework on gradual scaling and capability evaluation.
Circo Studio, Idea Lab — Academic series: Digital Symbiosis, Programming 3.0, Organizations 3.0. Internal documents, 2025.

Autonomy is delegated; never assumed. This document exists so that delegation is — every time — a debatable, reversible decision with a name attached to it.

Circo Studio Idea Lab Buenos Aires · Argentina · MMXXVI circostudio.io

Document open to revision.

ARL Manifesto · v1.1
Sheet 2 / 2
Circo Studio