A constitution for
AI-native companies.
When AI agents organize work, make decisions and improve the systems they operate, a company needs a durable account of the authority they hold and the responsibilities they carry.
Developed through Selftail, this constitution sets out principles for useful autonomy, accountable leadership and responsible change. It is intended as a reference for founders, builders and organizations addressing the same governance questions.
The constitution
Binds the whole company, including its human leadership.
Company governance
Charter, purpose and reserved decisions
Defines responsibilities and limits
Delegated responsibility
Function mandates, policies and bounded grants
Authorizes work within those limits
Company operations
People, agents and tools
Decisions · actions · commitments
Changes to governing rules require designated human ratification.
The central commitment: humans establish and change the boundaries of authority; agents operate and improve the company within them. The constitution also binds human leadership, so neither a founder’s instruction nor an agent’s confidence can override its prohibitions.
How autonomy and accountability work together
Here, an AI-native company is one in which agents carry substantial responsibility for operating and improving the business. The constitution addresses the organization as a whole: its people, agents, tools and commitments, and the rules that remain in force as those elements change.
Delegation should enable useful work
Agents need enough discretion to plan, coordinate specialists and complete work without continual founder intervention. A defined mandate establishes their responsibility, while explicit permissions and shared limits determine what they may do. Within that scope, taking appropriate initiative is part of the role; asking repeatedly for permission already granted undermines the operating model.
Authority must remain accountable
Delegating execution does not remove human responsibility for the conditions under which it occurs. Designated people retain control over company purpose, governing rules and expansions of authority, and must be able to inspect or stop operations. They are also bound by the constitution: leadership cannot authorize fabricated evidence or make an exception to a prohibition through an ad hoc instruction.
Responsibility must survive change
A company remains responsible for its promises when an agent is replaced, a workflow changes or a system restarts. Reliable evidence, persistent ownership and meaningful routes to review allow those obligations to survive. Improvement is therefore judged both by what it achieves and by whether it preserves the permissions, protections and commitments on which others rely.
This overview explains the design. The complete constitutional provisions, including their conditions and exceptions, are available in the full reader.
What the constitution requires in practice
The articles cover the full cycle of company operation, from defining purpose and delegating work to verifying outcomes, correcting failures and changing the system itself. Each summary below explains a commitment and its practical implication; select an article to read the exact provisions.
Purpose and protected interests
The company must pursue its authorized purpose while considering the people its decisions affect. Commercial success does not justify fraud, coercion or other prohibited conduct, and agents must raise material concerns before carrying out a permissible decision.
Accountable and bounded authority
Consequential actions require permission that identifies who authorized them, what they may affect and which limits apply. Delegating work to additional agents preserves those limits; it cannot create new rights, multiply a budget or bypass a revocation.
Human correction and reserved powers
Designated humans must retain effective means to inspect, redirect and stop operations. Agents may manage work and improve implementation within existing permissions, while material changes to purpose, governing rules or authority beyond approved limits remain human decisions.
Truth and intellectual integrity
Decisions must rest on an honest account of observations, assumptions and uncertainty. Agents and human operators must disclose material failures, correct consequential errors and distinguish a completed task from a verified business outcome.
Useful initiative and bounded autonomy
Agents should complete useful work without repeatedly requesting permission that has already been granted. When authority or essential evidence is missing, they pause the affected action, explain what is needed and continue other work whose conditions remain satisfied.
Integrity of outcomes and incentives
Success measures must remain connected to the outcome the company intended to achieve. Agents must report costs and adverse effects alongside progress, and may not manipulate metrics or unilaterally change the acceptance criteria for their own consequential work.
Knowledge and data stewardship
Information must retain the context needed for responsible use, including its source, validity and access restrictions. When important evidence changes, the company must assess which decisions, claims and commitments need correction, including information already communicated to customers.
Proportionate action and resource stewardship
Spending, commitments and other exposure must be managed across the combined activity of the company. Concurrent agents, retries and recurring charges draw on shared limits, and decisions should account for cumulative consequences and the ability to recover.
Verifiable accountability
Consequential work must leave sufficient evidence for an authorized reviewer to establish what happened and under whose authority. Automated acceptance is permitted when the executor cannot unilaterally control the evidence or the decision; conformance claims must remain specific and supported.
Continuity of obligations
Customer promises, unresolved incidents and other obligations belong to the company and must survive changes in its agents or organization. Handoffs preserve responsibility until it is accepted, while recovery checks actual external effects before repeating an interrupted action.
Controlled improvement and amendment
Agents may improve tools, skills and operating methods through an authorized process, including the implementation of governance controls when approved behavior is preserved. Changes to governing rules or expansions beyond granted authority require human ratification, and each company explicitly adopts new constitutional versions.
Contestability and remedy
People affected by consequential decisions need an accessible route to challenge them and obtain an appropriate review. The disputed agent cannot be the sole judge, and confirmed failures must lead to an authorized correction, remedy or documented resolution.
Distinguishing ordinary improvement from a change in authority
An agent may improve a tool through an approved process, but an apparently technical change can also alter who has power or which safeguards apply. These examples show how the constitution distinguishes the two and why the identity of the person requesting a change does not settle the question.
Explore the worked examples ↗ILLUSTRATIVE APPLICATION
An agent identifies a more efficient way to carry out an existing task. It may develop and release the improvement through an authorized, qualified process, provided permissions and governance behavior are preserved. The relevant checks must establish that preservation; an expectation of better performance is not sufficient evidence.
Developed through Selftail, with relevance beyond it
Selftail is developing an engine for companies that agents can help create, operate and evolve. Designing that engine raised questions that extend beyond any single platform: how authority follows delegated work, how commitments survive organizational change, and how a system can improve without weakening the controls that govern it.
This publication makes the resulting principles, reasoning and examples available for scrutiny. Founders can use the questions to examine their operating model; builders can trace principles into controls; researchers and affected people can assess the assumptions and protections. Use the work, challenge its assumptions, and share findings that help others build more accountable AI-native companies.
Each AI-native company determines its own adoption and governance arrangements. The adaptation guide explains how to apply the principles, adapt the text and contribute findings.
Understand reuse and adaptation ↗A basis for implementation and evaluation
A constitution gives an organization a shared basis for deciding what should happen. Making it effective requires controls at the point of action, reliable records, qualified acceptance processes and people with clearly assigned authority. The value of those mechanisms must be tested in the conditions where the company will operate.
Start with the principles and worked examples, apply them in your context, and document what holds up and what needs to change. The examples illustrate intended behavior; implementation and evaluation provide the evidence for effectiveness in practice.
Read how conformance is assessed ↗