The KYE™ Governance Spine™ decides every step of an agent run before it happens
When your agent moves money, reads a customer file, or hands work to another agent, each of those is a separate step — and each gets its own verdict before it executes. You get one thing auditors ask for and agent frameworks cannot give: proof of who authorised a specific action, at the moment it happened. Exam preparation drops from days of log reconstruction to minutes of replay.
The KYE™ Governance Spine™: side by side
Demo dataYou learn: How the options differ, criterion by criterion. You can: Pick a column to light it.
| Agent moves funds | A principal holding a purpose-bound grant | Two-person for irreversible amounts | Signed Evidence Pack™, verifiable offline | Revoke the grant; every later step freezes |
|---|---|---|---|---|
| Agent reads a customer record | A principal scoped to that tenant only | None; refused outright if cross-tenant | Access recorded with the purpose it served | Withdraw purpose; future reads deny |
| Agent hands work to another agent | The receiving agent, with narrowed authority | Inherited from the delegating grant | Delegation chain extends, never restarts | Revoke upstream; the whole chain falls |
| Counterparty changes in the real world | No one — the anchor moved, not the agent | Re-coupling before the next commit | Drift event mapped to an authority decision | Suspend or revoke, by declared severity |
How this widget is built
- Demonstrates
- Comparison
- Components
- The options this page compares
- The criteria it compares them on
- Flow
- Each criterion across the options, as the page sets it out. Architecture diagram
- Used on
- /governance-spine/: On this page, after its opening.
Run a governed pilot See the authority stack
Sample run. Each step is admitted before it executes; the refused tool call freezes everything downstream of it.
Eleven kinds of step, one decision point
A tool call is one kind of step, not the whole surface. If your agent is governed at its tool calls but not at its own model calls, boundary crossings or data reads, it is governed at some steps and ungoverned at others — and to anyone auditing the chain that reads the same as ungoverned.
- boot
- generation
- tool call
- API call
- access
- hop
- iteration
- resolve
- reconcile
- handoff
- terminate
Bold kinds are consequential — they change something outside the agent, so the verdict is sealed before execution. The rest are governed too, but change no state in the world.
What the spine controls, and what needs a person
Every capability below answers the same eight questions, because a governance claim you cannot interrogate is a badge. If you are a CISO evaluating this, the useful column is the last one: what happens when an action was wrong.
| Capability | Who may act | Approval | Audit output | Rollback |
|---|---|---|---|---|
| Agent moves funds | A principal holding a purpose-bound grant | Two-person for irreversible amounts | Signed Evidence Pack™, verifiable offline | Revoke the grant; every later step freezes |
| Agent reads a customer record | A principal scoped to that tenant only | None; refused outright if cross-tenant | Access recorded with the purpose it served | Withdraw purpose; future reads deny |
| Agent hands work to another agent | The receiving agent, with narrowed authority | Inherited from the delegating grant | Delegation chain extends, never restarts | Revoke upstream; the whole chain falls |
| Counterparty changes in the real world | No one — the anchor moved, not the agent | Re-coupling before the next commit | Drift event mapped to an authority decision | Suspend or revoke, by declared severity |
Where the spine stops
Most governance pages tell you what a system covers. The useful half is the opposite, and these are the limits your auditor will find anyway — so we would rather you heard them from us.
- Revocation is bounded, not instantaneous. A step already admitted and running completes. Admission is a lease with a measured expiry, so the exposure is one lease interval — a number in the contract instead of a surprise in an audit. We do not claim a zero window; physics is indifferent to constitutions.
- KYE Protocol™ does not run your agent. Your runtime does the thinking. KYE™ governs the authority to act, and keeps the non-competition line with agent frameworks explicit.
- Coverage is measured and published, including the gaps. Of 174 lifecycle machines in the estate, the ones that declare full transition sets are counted separately from the ones that do not — because a machine that cannot say whether a state can be left should not be reported as covered.
- An unmapped regulation stays unmapped. KYE Protocol™ will not claim enforcement of a framework it has not mapped to a real artefact. Out-of-scope is an honest state; inflating coverage is not.
Evidence you can check without us
Every verdict seals into a portable record: what was decided, by whom, under which grant, against what the world looked like at that instant. Your auditor verifies it from published keys alone — no KYE Protocol™ account, no API call, no network.
That matters most exactly when it is hardest. A governance layer that must phone home to reach a verdict fails when the network fails, which is when authority certainty matters most. Decisions resolve locally, and so does verification. See the Evidence Pack™ and air-gapped authority surfaces for how the record is built and checked.
Honest scope: this is an audit record with a verifiable signature, not a legal opinion. It proves an action met the declared controls at runtime; it does not decide whether those controls were the right ones.
Canonical KYE™ surfaces referenced on this page: KYE Protocol™.