The boundary between permission and consequence.
AI agents are being given access to APIs, payments, email and code execution. What most deployments lack is a record, checkable by anyone, of who authorised each action, for what purpose and within what limit. Eden Control records who authorised each action an agent takes through it, checks the action against that authority before it runs, and holds or refuses the rest.
A working prototype: it governs its founder’s own AI agents today.
- Principal
- Finance team
- Agent
- Payments agent
- Action
- Pay supplier invoice 2207
- Purpose
- Monthly supplier payment
- Budget
- Up to £4,800
- Expiry
- Today, 17:00
- Identity of the principal verified
- Amount within the delegated budget
- New supplier: routed to the finance controller
Three reasons this cannot wait.
AI agents have no seatbelt
Companies are racing to give AI agents access to production systems, bank accounts, email, and code execution. Few of those deployments keep a record, checkable by anyone, of what was authorised, by whom and within what limits.
Liability evaporates without a record
When an AI agent makes a consequential mistake, the question is always: who authorised this? Without an authority record, the answer is nobody, and everybody. Organisations need durable, independently verifiable proof of who decided what.
The law already asks for records
The EU AI Act requires high-risk AI systems to keep logs of their operation and to allow effective human oversight. Duties like these are mostly met with paperwork today. Eden Control makes the record part of the action itself.
AI proposes. Authority decides. Evidence survives.
In Eden Control, every consequential AI action requires an authority envelope before it can execute. The system checks the envelope against deterministic patterns. If clear: Allow. If uncertain: Investigate, routed to a human reviewer with an expiry. If blocked: Deny. It fails closed, with no exceptions. Every outcome produces a cryptographic receipt. No valid authority, no action.
Exact-action authority
Bind principal, agent, action, purpose, budget and expiry into a cryptographic authority envelope. No envelope, no action.
Contestable Investigate
Route uncertainty to named human reviewers with evidence, expiry, dissent and a fail-closed timeout. Hesitation is the control.
Portable outcome proof
Authority, decision, execution and outcome connected in a single receipt, designed to be verified offline by anyone, with no trust in the system required. Under test: hardening for external audit is the active programme.
Built to govern the agents that built Eden.
Eden Control began as the control layer its founder built to govern the AI agents that were building Eden Legal AI: every agent action checked against its rules and allowed, held for a person, or refused before it ran. The same system runs Eden Research’s checks. Its register held 1,244 written rules on 14 October 2026, 178 of them with an automatic check of their own; between 5 July and 14 October 2026 those checks returned 1,514,217 decisions, holding 4,877 actions for a person and refusing 968.
From a book to a working system
The idea comes from the founder’s book, Infinite Architects (January 2026), and the Eden Protocol: correction built into the loop from the start, not added afterwards. Eden Control is that idea applied to the actions of AI agents.
Three answers, not two
A check with only two answers must pass or fail what it cannot decide. A third answer lets uncertainty wait for a person, so nothing slips through on doubt.
One system, three disciplines
Eden Control governs the actions of AI agents, Eden Legal AI applies the same discipline to litigation, and Eden Research to science.
Clear status beats inflated claims.
We publish what is built, what is under test, and what remains a release gate.
Operator shell
Separate Control shell with explicit tenant context, routed gateway domain, and Google identity verification at the API boundary.
Ternary gate engine
A register of 1,244 written rules across legal, development, operations and shared work (counted on 14 October 2026), with about 176 automatic checks run before an agent’s action. From 5 July to 14 October 2026 they returned 1,514,217 decisions, holding 4,877 for a person and refusing 968. Uncertainty holds.
Portable authority envelope
Exact-action delegation with offline-verifiable outcome receipts. Hardening for external audit is the active programme.
External assurance
Independent penetration testing and published bypass-resistance evidence remain explicit release gates. We will not claim security we have not proven.
What people ask first.
What is Eden Control?
A governance layer for AI agents. It records who authorised each action an agent takes through it, checks the action against that authority before it runs, and holds or refuses the rest. It is a working prototype that governs its founder’s own AI agents today.
What is an authority envelope?
A record that binds the principal, the agent, the action, its purpose, its budget and its expiry before the action can run. No envelope, no action.
What happens when a check is uncertain?
The action is held and routed to a named human reviewer, with the evidence and an expiry. If no one decides in time, it fails closed.
What is implemented today?
The operator shell and the gate engine are implemented. The portable authority envelope and external assurance are under test. Eden Control claims no regulator approval or external certification.
Who built Eden Control?
Michael Darius Eastwood, who built it to govern the AI agents that were building Eden Legal AI. The same system checks the claims of Eden Research.
One system. Three disciplines. The same standard.
Can correction keep pace with an improving AI?
The science behind the method: registered predictions, stated before the tests that could refute them.
Visit Eden Research → Eden Legal AILitigation intelligence that shows its work.
One linked record for the whole case, every authority checked at its official source, and a person’s approval on everything court-facing.
Visit Eden Legal AI → The founderMichael Darius Eastwood
The book, the research programme, and the record behind all of it.
Visit his website →