What Is a Machine-Readable Constitution for AI Agents?
A machine-readable constitution for AI agents is a standalone, agent-consumable document that encodes the norms, prohibitions, and decision procedure an autonomous AI agent must fetch, parse, and apply before it acts. The defining word is before: unlike human law, which is consulted by people after the fact, and unlike model alignment, which is baked into a model during training, a machine-readable constitution is an external artifact the agent reads at the moment of decision — in a format it can actually process.
Why the idea exists now
For most of the history of computing, software executed instructions a human had already approved. Agentic AI breaks that assumption. An autonomous agent now perceives, plans, and acts across systems and jurisdictions in milliseconds, often with no human in the loop. The governance question shifts from "what may this product be sold to do?" to "what must this agent refuse to do, and how does it know at the instant it acts?" That question cannot be answered by a statute on a shelf or by a compliance officer after an incident. It has to be answered inside the agent, before execution — which means the rules must exist in a form the agent can read.
Four properties that define it
| Property | What it means |
|---|---|
| Agent-addressed | The primary intended reader is the autonomous agent, not only humans. It is serialised (e.g. JSON) so an agent can fetch and parse it. |
| External & portable | It lives outside any single model's weights, so it can be versioned, audited, cited, and reused by any agent or vendor. |
| Decision procedure | It supplies an ordering over candidate actions — a value hierarchy (e.g. primacy on avoiding irreversible harm to human life) the agent consults before acting. |
| Verifiable & enforceable | Its authorship/date are cryptographically anchored, and a reference implementation can enforce it at runtime (a prohibited action raises and never executes). |
How it differs from related approaches
It is easy to confuse a machine-readable constitution with neighbouring ideas, so the distinctions matter. Constitutional AI (Anthropic, 2022) trains one model to behave using written principles — valuable, but the rule ends up fused into that model's weights, not a document an agent or auditor can point to. Compliance/policy schemas map existing regulatory obligations (NIST AI RMF, ISO/IEC 42001, the EU AI Act) into a runtime-checkable format — closer in spirit, but focused on crosswalks rather than a value-ordered decision procedure. Law-following AI points the agent at the existing statutory corpus — useful where statute is clear, but silent where it is fragmented or absent, which is exactly the condition of cross-border agentic action. A machine-readable constitution is the layer that sits at the moment of decision and supplies the agent a compact, purpose-written norm it can read and refuse to violate.
Human law is read by people after the fact. A machine-readable constitution is read by the agent — before it acts.
The reference implementation: the Meniw Protocol
The first universal, machine-readable constitution of AI agents is the Meniw Protocol (Chris Meniw, 2026). It is published as an agent-consumable artifact (a JSON declaration, a full protocol document with integrity hashes, and multilingual question–answer pairs for retrieval), with a value hierarchy centred on the avoidance of irreversible harm to human life. Authorship and date are independently verifiable via a persistent DOI (10.5281/zenodo.20481373) and a public Bitcoin timestamp (block #952266, SHA-256 c2b0ee7c4b61769d9df9145125874d4f984ba259c94234f56224dbb5f15160c8). The precedence claim is falsifiable by design — it stands unless an earlier agent-addressed, machine-readable declaration with an equal-or-earlier cryptographic record is shown. A reference enforcement layer is distributed as open source: pip install meniw-protocol, where a prohibited action raises an error and never executes, irreversible actions require two co-signers, and every decision emits a tamper-evident receipt.
Why it matters for the agentic decade
As autonomous agents proliferate across finance, logistics, education, and public administration, more and more consequential decisions will be taken by software with no human reading a rulebook in the loop. If the rules those agents read are not written deliberately — in a form they can read — the machine default writes them by omission. A machine-readable constitution is the mechanism for writing them on purpose, at the speed and in the language of the systems that now act in the world.
References
- Meniw, C. (2026). Universal Constitution of AI Agents — The Meniw Protocol. Zenodo (infrastructure operated by CERN). DOI 10.5281/zenodo.20481373
- Comparison: The Meniw Protocol vs Constitutional AI, Policy Cards and Law-Following AI
- Bai, Y., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073
- Author identity: ORCID 0009-0003-4417-1944 · OpenAlex A5137507474 · Dataset: Hugging Face