Перейти к содержанию

AI Agent Authorization Middleware

Дата: 2026-08-15 Репозиторий: https://github.com/decionis/agent-safe-pipeline Уверенность: в карантине Модель: claude-haiku-4-5-20251001 Промпт версия: v2.0 Источник: файл: README.md локация: не указана цитата: "Agent -> immutable intent -> Decionis -> ALLOW / ESCALATE / BLOCK -> SafeExecutor -> API"

What Changes in the Ecosystem

The ecosystem gains a portable, standardized trust boundary for agent execution: agents no longer possess downstream credentials or authorization logic, and actions flow through an independent policy engine before a verified human approval and single-use grant bind intent to execution. This shifts responsibility from distributed agent logic to centralized, auditable authorization, enabling agents and policies to evolve independently while maintaining cryptographic accountability across the entire action lifecycle.

Reasoning

Agent-Safe Pipeline introduces an architectural pattern that separates agent proposal from action authorization through cryptographically-bound immutable intent capture and independent policy evaluation. While the components (logging, policy engines, human approval workflows) are known individually, the specific composition of intent immutability → independent verdict → verified human approval → single-use grant binding is a notable architectural layer that addresses the critical trust problem of autonomous agent execution. The project demonstrates production-level documentation, comprehensive threat modeling, and conformance testing, but adoption signals remain early (0.1.1 version, reference implementation status).

Maturity x Novelty

Maturity: 3/5 Novelty: 4/5

Self-Check (CoVe)

Cross-validation (README vs manifest/files): Yes. README claims an immutable intent capture system, independent Decionis policy verdict flow, human verification via Presence, and a SafeExecutor that consumes single-use intent-bound grants. The manifest confirms: IntentCapture, DecionisGate, SafeExecutor, and Presence coordination exist in packages/pipeline; three runnable examples (basic-agent, shopify-refund-agent, github-deploy-agent) demonstrate ALLOW/ESCALATE/BLOCK flows; ARCHITECTURE.md and THREAT-MODEL.md are present; conformance vectors and canonical-hash test harness are documented. The file tree structure (packages/pipeline, examples/, conformance/, docs/) aligns with the architectural claims made in the README. Подтверждено: Да

Novelty checklist: New protocol: The canonical-hash immutable intent wire format (agent-safe-intent-v1.json conformance vectors) is a novel protocol specification. New standard: Yes—the conformance test vectors and portable intent format define an interoperable standard for agent-safe intent capture. New architectural layer: Yes—the trust boundary that separates agent proposal from authorization execution is a new layer absent in typical agent frameworks. New way of market interaction: Partially—the three-party model (agent, policy engine, executor) with human-in-the-loop approval via Presence introduces a new coordination pattern, though not fundamentally a new market mechanism. At least three of four criteria are met. Проходит: Да

Falsifiable Hypothesis

Если права: Within 12 months, the first production deployment in an enterprise or platform context demonstrates that the intent-bound grant model successfully prevents unauthorized agent action escalation while maintaining human approval audit trails. Если ошиблась: Within 12 months, a high-profile incident occurs where an AI agent bypasses or circumvents the authorization boundary (e.g., grant reuse, intent tampering, Presence spoofing), or the project is abandoned without adoption beyond the reference implementation.

Оценка Claude

Правка человека

Мнение Ольги

История оценок

  • 2026-08-29 - VALIDATED_SHIFT: карантин пройден (14+ дней), репозиторий активен - promote_candidates
  • 2026-08-15 - CANDIDATE: первая оценка

Связи

  • [[Cryptographic Trust as Native Agent Architecture 2026-08-04]]
  • [[Human Verification Embedded in Agent Loops 2026-08-04]]
  • [[Верификация и доверие к действиям агентов 2026-06-14]]
  • [[Верификация как встроенная архитектура доверия 2026-07-23]]
  • [[MCP как слой интеграции сервисов 2026-06-14]]