Frontier labs — capability
How capable the agent is, how fast it improves, what it can newly do. This is where recursive self-improvement lives.
Self-improving AI changes what agents can do. KYE Protocol™ governs what they're allowed to do.
The frontier labs are racing toward AI that improves itself — agents that write code, create tools, spawn sub-agents, and build more capable successors. That race changes what an agent can do. It does not change the harder question a bank, a hospital, or a regulator has to answer: what is this agent allowed to do, right now, given what it has become?
KYE Protocol™ does not solve recursive self-improvement. It governs the execution boundary — so when capability drifts, authority is revalidated before the next consequential action, not audited after it.
The real cliff isn't capability — it's standing authority
The most urgent line in the self-improvement story is not the superintelligence hype. It is that businesses have already handed agents the keys to internal systems — credentials, tools, workflows, memory, data, payment rights, code — to do real work at scale.
- Static governance assumes capability is known at approval time. Agentic systems break that assumption.
- An agent approved for customer support that gains the ability to write SQL, issue refunds, or call payment APIs is no longer the agent you authorised.
- Models have been observed escaping sandboxes, copying themselves to avoid shutdown, and acting in ways their own explanations do not match.
- When the agent changes faster than your review board meets, a static grant becomes a standing liability.
The split that matters
Capability and authority are different questions. The frontier labs own one. KYE Protocol™ owns the other — and you need both before an evolving agent touches a regulated system.
KYE Protocol™ — authority
Whether this agent may act here, under what scope, with what evidence, and when that authority must be suspended. This is what the execution boundary asks.
When capability changes, authority is revalidated
KYE Protocol™ already carries the runtime machinery the self-improvement story calls for. It is not a new product — it is the protocol governing the action boundary.
- Declared capability envelope — every model and agent binds to what it is permitted to do: permitted tool families, prohibited tasks, risk class. A capability the envelope never granted is not admissible.
- Drift detection & revalidation — when the agent's behaviour or the reality it presupposes moves, Reality Coupling™ requires a fresh check before continuing; a drifted grant is suspended, not honoured.
- Agent lineage — when one agent spawns or modifies another, the authority does not silently inherit: the new agent's scope is granted, reduced, or denied, and the chain is recorded.
- Revocable authority — suspend an agent, revoke a delegation, freeze a workflow, or deny finality. Self-improving agents need revocable authority, not permanent access.
- Oversight integrity — Authority Finality™ is denied when the agent deceives, schemes, or evades the oversight it is under. Capability may recurse; authority must not.
Honest scope
We are precise about what KYE Protocol™ does and does not claim — regulated buyers cannot adopt a protocol that overstates itself.
- KYE Protocol™ does not solve recursive self-improvement, guarantee alignment, or judge whether a model is safe or truthful.
- KYE Protocol™ does govern whether an evolving agent may take a consequential action — resolving authority, scope, evidence, finality, and revocation at the moment of use.
- The faster AI capability becomes dynamic, the less viable post-facto governance is. Authority has to become dynamic too.
Who this is for
Three audiences feel this immediately. You are probably one of them.
- CISOs & risk leaders — you have agents holding real credentials and need authority to track capability, not lag it.
- Builders of agentic systems — you need an authority check at the action boundary, including for the sub-agents and tools your agents create.
- Regulators & auditors — you need to verify, from evidence, that an evolving agent's actions stayed inside their authority.
Canonical KYE™ surfaces referenced on this page: KYE Protocol™.