2AYE
← All articles

Architecture and verification

AI agent authorization, from human intent to execution evidence

How 2AYE connects identity, signed intent, runtime policy, human approval and execution evidence, with verified results and remaining release gates.

By 2AYE ·

Evidence and availability: this article describes the architecture and dated verification results. Local tests, a deployed website and a signed app build establish different things. They do not establish complete customer acceptance or general production readiness. Check current capability status before choosing a deployment.

The gap between access and action

An agent can have a valid identity and still propose the wrong refund, export too many records or change a resource outside its remit. Sign-in is one boundary. The consequential action needs its own enforceable authority.

2AYE is built around a signed intent: a human-defined purpose with deterministic limits. An agent proposes an action; the platform evaluates its exact parameters, requests human approval when needed and issues a short-lived, single-use grant. Enforcement at the protected resource and an execution receipt complete the loop.

Existing identity providers, business systems and AI models remain part of that architecture. Integration requires evidence at each boundary, including the destination that can actually make the change.

The end-to-end authorization loop

  1. 1. Identify the caller

    Establish the person, workload or agent and its tenant. A successful sign-in identifies the caller; action authority is evaluated separately.

    Explore this control
  2. 2. Sign the intent

    Record purpose and deterministic limits in a versioned, signed contract. Delegation may narrow authority; changing material terms requires a new signature.

    Explore this control
  3. 3. Evaluate the exact action

    Check the registered action parameters against the contract, current policy and available accounting. Deterministic errors deny; semantic review can escalate but cannot override a denial.

    Explore this control
  4. 4. Escalate when approval is required

    Present the exact action for a device-signed human decision. An approval applies to that action, rather than giving the agent general permission.

    Explore this control
  5. 5. Enforce at the destination

    A short-lived grant binds the evaluated parameters and may be redeemed once. A protected receiver must refuse altered parameters, expired authority, replay and alternate mutation paths.

    Explore this control
  6. 6. Record and reconcile the outcome

    Keep the decision, redemption, execution receipt and audit evidence connected. If an external outcome is uncertain, reconcile it before attempting another write.

    Explore this control

Find the capability behind each step

Six product families and shared capabilities organize the platform. The links below explain their scope; they are not a claim that every integration or deployment model is available. The evidence-based inventory and full site directory provide the broader map.

Agent
Agent identity, ownership and governed lifecycle.
Authorize
Signed intent, deterministic policy and exact-action grants.
Approve
Human review and device-signed decisions.
Auth
Sign-in, sessions and supporting identity factors.
Risk
Risk findings and escalation within the authorization boundary.
Evidence
Receipts, audit and portable verification artifacts.
Policy and delegation
Constraints, narrowing and authority changes.
Integrations
APIs, SDKs, MCP and receiver integration boundaries.
Verification
How your team can check the permissions and approvals 2AYE signs.
Recovery
Reconciliation and operational recovery requirements.
Mobile
Authenticator and approval app scope and release status.
Deployment
Current controlled deployment and planned deployment models.

How we check it

Every release is tested against the controls above, including refusal of changed parameters, expired permissions and reused permissions. Test results describe 2AYE's own testing, not your deployment or an independent validation.

Architecture, test evidence and verification material are shared with evaluating customers and partners in a security review.

What is not available yet

  • Organizations are onboarded by invitation while organization review is completed.
  • The 2AYE Authenticator is not yet available in the app stores, and approvals on physical iPhone and Android devices are still being completed.
  • Live payments are not available. Connections to your providers and systems are set up and checked with each customer.
  • Customer-managed signing keys are not currently offered.
  • Private-key client authentication, DPoP and agent identity federation are planned, not yet available.

No government authorization, validated-cryptography certification or Okta agent interoperability is claimed here. Review the security boundary and deployment availability for the intended environment.

Start with one consequential workflow

Choose a bounded action and identify the system that performs it. Define who may authorize it, the exact parameters that matter, how uncertainty is escalated and how the destination refuses bypasses. Test denial, tampering, expiry and replay alongside the intended successful action. Keep a receipt that can be checked after execution.

Use the developer quickstarts, API reference and failure-state guide to assess the implementation boundary. For a controlled evaluation, discuss your workflow with 2AYE.

Share this article

Use your device share menu for Messenger and other installed apps. Available apps depend on your device. You choose the recipient and confirm the message.