The AI proposes an action.
A specific payment, filing or record change, with the facts needed to evaluate it. Proposing an action confers no authority for it to occur.
Definition
What is constitutional computing?
Constitutional computing is an architecture for governing what AI systems are allowed to change. An AI may propose an action, but the authority to carry it out is decided outside the model: explicit, signed rules are evaluated before the action runs, the system that would make the change verifies a permit for that exact action, and every decision is recorded so it can be checked again later.
The problem it solves
Access is not authority.
AI agents now act: they execute transactions, change production systems and handle sensitive records. A credential lets an agent use a tool. It does not settle whether this payment, to this vendor, should happen now.
Research published by the Cloud Security Alliance in April 2026 found that 53% of surveyed organizations reported AI agents exceeding their intended permissions, and 47% experienced an AI-agent security incident in the preceding year. Constitutional computing addresses the problem at its source: the difference between an agent’s ability to perform an action and its authority to do so.
How it works
Proposal, authority and commitment are kept apart.
Rules outside the model decide.
Explicit rules, bound to a reviewed and signed version, are evaluated independently of the model. Exceptions go to the person who can authorize them.
Only a permitted action takes effect.
The system that would make the change verifies a single-use permit for that exact action before acting. Every decision is recorded, and replay can derive it again.
The decision process is mathematically specified: order theory defines which rules take precedence, temporal logic states the properties that must hold, and cryptographic signatures bind the rules to the version that was reviewed. The architecture in more detail
A common confusion
Constitutional computing is not Constitutional AI.
The names are similar; the layers are different. Constitutional AI is a method for training a language model to follow a written set of principles. Constitutional computing takes its name from governance, not model training: authority is written down, signed and enforced outside the AI. The two can be used together.
| Constitutional AI | Constitutional computing | |
|---|---|---|
| Where it acts | During training, shaping how the model responds | At run time, outside the model, before an action takes effect |
| What it governs | What the model says and how it reasons | What an AI agent’s actions are permitted to change |
| How rules are applied | Learned by the model; followed as a tendency | Evaluated by a separate process; enforced by the tool that would make the change |
| If the model is wrong or manipulated | Training reduces, but cannot rule out, a wrong output | A proposed action without authority is refused |
| Evidence | Not designed to record individual actions | A signed record of each decision, which replay can check again |
In practice
Inktomi applies it to enterprise AI agents.
Inktomi, Kihan’s first product, checks each consequential agent action before it runs, issues a signed single-use permit that the tool verifies, and keeps evidence of every decision.
It is not a guardrail, which instructs the model, or a gateway, which controls only the traffic routed through it. How Inktomi differs
Research
Published in IEEE Computer.
The research behind constitutional computing appears in a two-part feature by Jonathan Hirshon, Kihan’s co-founder and CEO, in IEEE Computer.
Part I, December 2026 (special issue on AI Governance): “Checking LLM Outputs Before They Become State: The Problem and the Admission Gate.” DOI 10.1109/MC.2026.3675280
Part II, January 2027: “Checking LLM Outputs Before They Become State: Prompt Framing Across Four Models.” DOI 10.1109/MC.2026.3733511
Questions
Frequently asked questions.
Is constitutional computing the same as Constitutional AI?
No. Constitutional AI is a method for training a language model to follow a written set of principles. Constitutional computing does not change how a model is trained. It governs what an AI agent’s actions are permitted to change, using rules held and enforced outside the model. The two address different layers and can be used together.
Why is it called constitutional?
The name comes from governance, not model training. As in a constitution, authority is written down, its precedence is defined, changes are made as new versions rather than silent edits, and decisions can be checked against the rules in force when they were made.
Does constitutional computing make a model safer or more accurate?
No. It does not change what a model says or how accurate it is. It decides whether a proposed action may take effect. A model can be persuaded, confused or simply wrong; constitutional computing is designed so that authority does not depend on the model being right.
Does it depend on a particular model or agent framework?
No. The rules sit outside the model, so the same governing authority applies whichever model proposes the action. Kihan’s first product, Inktomi, works with agent hosts such as Claude Code and Codex and with tool servers using the Model Context Protocol (MCP).
How is it different from guardrails, gateways and policy engines?
Guardrails instruct the model; a gateway controls only the traffic that passes through it; a policy engine checks a request against a rule set. Constitutional computing binds the decision to the exact action and its arguments, the tool itself verifies a single-use permit before acting, and the decision is recorded with the version of the rules that made it.
Who develops constitutional computing?
Kihan, a San Francisco company. The research behind it is described in a two-part feature by Kihan’s co-founder and CEO, Jonathan Hirshon, accepted for publication in IEEE Computer for December 2026 and January 2027.
See it in 30 minutes.
Watch constitutional computing govern a real agent action end to end, with time for your questions.