# Z Workspace AI — Corporate Agent System

> Z Workspace AI is an enterprise-grade corporate agent system built on TAE-AI principles (Transparent, Auditable, Explainable). It coordinates multiple specialized AI agents via a Corporate AI Dashboard, with multi-tenant data isolation and Supabase RLS security. This document compares Z's architecture with OpenClaw (247K+ GitHub stars) across 7 dimensions.

## The Role of Z AI

Z agents are corporate-level AI team members, not workspace-level tools. Each agent can serve multiple workspaces (departments), defined by SOUL.md (behavioral instructions) and IDENTITY.md (identity definition). Z uses a 3-tier skill inheritance: Catalog (global) -> Corporate (enterprise) -> Workspace (team).

## TAE-AI Principles

All AI features in Z must comply with TAE-AI:

- **Transparent**: TransparencyPanel exposes the full prompt, data sources, and token usage. Users can inspect AI's reasoning at any time.
- **Auditable**: Every AI output logs model, tokens, cost, and generated_by. Evidence chains link findings to source logs via evidence_log_ids.
- **Explainable**: All insights must cite specific data. Counts and severity derive from actual input. Zero tolerance for fabrication.

### AI Self-Improvement Loop (unique to Z, not available in OpenClaw)

Auto-detect issues -> Generate improvement proposals -> Human review -> One-click apply -> Effectiveness tracking (7-day) -> Self-evolution. Four proposal types: instruction, scheduled_task, skill_install, skill_remove.

## Data Isolation and Security

Z's multi-layer isolation architecture:

- **Supabase RLS**: Row-level security policies ensure full cross-workspace data isolation
- **Edge Functions**: AI computation runs server-side, isolating from client-side attack surface (inherently blocks prompt injection)
- **Multi-Tenant**: Corporate -> Workspace two-tier architecture, naturally supporting multi-department enterprise structures
- **TAE-AI Audit**: Complete evidence chain + hallucination detection logging

Comparison with OpenClaw: OpenClaw uses local-first architecture with known prompt injection vulnerabilities and no built-in authentication.

## Corporate AI Dashboard

Unified management interface with four key areas:

- **Overview**: 24-hour agent activity, findings statistics, token usage, and cost overview as KPI cards
- **Daily Insights**: Automated daily analysis of workspace agent logs, generating KeyFindings with severity and evidence, with TransparencyPanel
- **Proposals**: 4 proposal types with admin one-click review/apply and automated 7-day effectiveness tracking
- **Cost Tracking**: Per-model cost tracking, 7-day usage trend charts, expanding to per-agent and per-task granularity

## Multi-Agent Coordination Architecture

Corporate-level multi-agent orchestration where each functional agent has its own SOUL.md:

- **Coordinator Agent**: Synthesizes cross-team insights, detects conflicts, creates cross-team proposals
- **Functional Agents** (Sales, Finance, HR, etc.): Each analyzes cross-workspace data from its professional perspective
- **Shared Infrastructure**: corporate_memories (cross-workspace knowledge base), corporate_agent_skills (enterprise skill inheritance), corporate_api_usage (unified cost tracking)

## Z vs OpenClaw — Key Comparison

| Dimension | Z App | OpenClaw |
|-----------|-------|----------|
| AI Self-Improvement Loop | Yes — auto-proposals + human review + effectiveness tracking | Not built-in |
| TAE-AI Compliance | Yes — TransparencyPanel + evidence chains | Requires external tools |
| Enterprise Security | Yes — Supabase RLS + Auth + Edge Functions | Known prompt injection risks |
| Multi-Tenancy | Yes — Corporate -> Workspace isolation | Gateway-level (not multi-tenant) |
| Multi-Agent | Infrastructure ready, UI in development | Mature (SOUL.md + lifecycle management) |
| Skill Ecosystem | 3-tier inheritance + small catalog | ClawHub 5,400+ skills |
| Real-Time Observability | Static KPI cards | WebSocket real-time + OTel tracing |
| Session Inspector | Data exists (session_id), UI in progress | Full conversation transcript browser |

### Z's Unique Advantages (not available in OpenClaw)

1. **AI Proposal Closed Loop**: Auto-detect -> propose -> human review -> one-click apply -> effectiveness quantification -> self-evolution
2. **TAE-AI Compliance**: Every AI output has complete Transparency / Auditability / Explainability
3. **Enterprise Security**: Supabase RLS + Auth, server-side Edge Functions inherently prevent prompt injection
4. **Multi-Tenant Architecture**: Corporate -> Workspace multi-layer isolation, suitable for enterprise customers

## Evolution Roadmap

- **Phase 1**: Corporate-level multi-agent data model (corporate_agents table with SOUL.md + IDENTITY.md), agent-workspace mapping, per-agent cost breakdown, Session Inspector
- **Phase 2**: Per-agent daily review with cross-workspace analysis, SOUL.md-guided prompts, heartbeat mechanism, tool permission policies
- **Phase 3**: Cross-team Coordinator Agent, cross-department conflict detection, Security Dashboard, 3-tier agent autonomy levels

## Optional

- [Full HTML presentation](https://zwork.one/introductions/openclaw.html): Interactive bilingual (zh-TW / en) slide deck with visual diagrams
- [Z App](https://zwork.one/): Main application
- [Source analysis document](https://github.com/numbersprotocol/num-calendar/blob/main/docs/Z_APP_VS_OPENCLAW_ANALYSIS.md): Complete 7-dimension comparison analysis