Orqetra: The Governance Layer for AI Agents. Enforce safe execution

Hi Product Hunt community! 👋 I’m Teppei, Founder of Utena Inc.

We built Orqetra because we realized a critical truth in the enterprise AI space: The missing layer between AI intelligence and enterprise adoption isn't better prompts—it's governed execution.

Today, LLMs are becoming incredibly smart. They can write, summarize, and plan. However, from a corporate security (CISO) perspective, there is a massive wall. No enterprise can blindly trust a black-box AI to control local operating systems, touch legacy admin screens, or execute destructive actions without strict guardrails. Truly difficult part is not "Can AI give a good answer?" but "Which screen should AI target, how far can it execute, and what should be verified?"

Orqetra is a Human-directed Execution OS built to solve this exact problem. Instead of letting AI run wild, we enforce a strict "Separation of Powers" framework:

- Human: Retains intent, judgment, and final approval.

- AI: Handles context understanding, composition, and quality evaluation.

- Program: Secures execution boundaries and handles local verification.

What makes Orqetra unique (and fundamentally different from standard agents):

1. Target Capture Lease (Privacy First)

We completely reject constant screen monitoring, as it is rarely accepted in corporate environments. Orqetra only observes the screen surface for a short lease window when explicitly triggered by the user (Detect current surface / Use current foreground surface), fully protecting corporate privacy.

2. Contract-Backed Architecture (6 Core Contracts)

Every AI action is bound by 6 rigid contracts within our execution kernel layer before reaching the real world:

- Runtime Input Contract: Treats URLs, file names, and destinations as temporary, runtime-only values.

- Browser Observation Contract: Limits retrieved data to safe, visible text and meta-information.

- AI Composition Contract: Restraints AI generation with clear goals, inputs, evidence, and formats.

- Output Quality Contract: Checks for insufficient evidence, missing fields, or placeholders.

- Human Approval Contract: Never proceeds with critical actions (save, send, write) without explicit human approval.

- Execution Verification Contract: Verifies post-action results (e.g., whether it was successfully saved or sent).

3. Legacy System & Desktop Compatibility

Orqetra does not rely solely on modern APIs. It is designed to govern actions seamlessly on old SaaS tools, browser-based internal portals, and native Windows environments where APIs do not exist.

4. Cost & Security Efficiency

Sending every click and UI change to external LLMs inflates API costs and risks privacy leaks. Orqetra's Browser Observation and Windows Agent handle observation, execution, and verification locally on the program side. We focus expensive LLM tokens exclusively on high-value reasoning.

5. Growing Through Safe Shared Structure

Concrete inputs, URLs, and DOM values remain runtime-only and are never saved as shared skills. Instead, Orqetra anonymizes and generalizes only the execution structures (workflow patterns, input schemas, quality gates, and contract definitions) into reusable skills across the organization.

AI thinks. Orqetra binds execution with evidence. Humans approve critical actions. Programs secure the boundaries.

Our core architecture is already patent-filed in Japan (Patent App: 2026-095248), and we are preparing for international PCT and US phases. While our full local runtime is in a Closed Alpha phase for security, we provide a secure demo path on Cloud Run to review our reference implementations for AI composition, quality gates, and contract design.

If you are building, investing in, or leading enterprise AI infrastructure, I would love to exchange perspectives. My DMs are always open! Let me know what you think. 👇