Public showcase · Governed AI Actions

General Agent Framework — Governed AI Actions

Build AI agents that can safely act on real business systems.

Human approval. Durable execution. UNKNOWN-aware reconciliation. Auditable external actions.

This public showcase contains documentation and demos. The commercial implementation remains private and separately licensed.

Primary business story

宏晟智造 · Hongsheng Works

A small factory moves from quotation to shipment. AI assistance proposes; approvals and role boundaries determine the next action.

PUBLIC SIMULATION MODE

Deterministic synthetic data and scripted AI assistance. No live ERP/MES or real customer; no commercial Framework execution proof. Human approval is a DEMO HUMAN APPROVAL SIMULATION.

Hongsheng Works owner dashboard showing pending approvals, overdue production and inventory risks in public simulation mode
Owner dashboard: pending approvals, overdue production and inventory risks. Chinese-first demo UI; synthetic data. View full-size screenshot.
Follow the quotation → approval → shipment story

Four control concerns

An action needs more than a good suggestion

01

Authorization & approval

Examine the proposal and bind the approval to the intended action. Keep requester and approver responsibilities separate.

02

Durable intent

Record an intent and execution attempt before an external call, so an interrupted operation leaves evidence.

03

Honest uncertainty

A lost response can leave an action UNKNOWN. An uncertain write must not be blindly repeated.

04

Evidence & audit

Keep decisions and observations inspectable. Approval-task chain verification has a specific scope.

Conceptual concerns illustrated by the technical refund demo; the manufacturing simulation shows business behavior and governance boundaries, not commercial execution.

Read about governed actions →

Response lost. Two different questions.

UNKNOWN can coexist with MATCHED

Command truth

UNKNOWN

We never received a confirmation for the command.

External reconciliation verdict

MATCHED (derived)

A fresh, usable observation agrees with the intended change.

Reconciliation does not turn the command into SUCCESS. These are separate axes, not two names for completion.

Understand the lost-response case →

Secondary technical demo

A governed refund with a lost response

The Framework-backed refund demo uses synthetic data and simulated human approval; it contacts no live business system. Re-running it requires the separately licensed private Framework. The manufacturing simulation runs independently.

The technical demo substitutes the transport while exercising the licensed Framework's execution behavior. That scope does not apply to the manufacturing public simulation.

Explore the technical scenario and its limits →

Read the concepts. Understand the boundary.

A conceptual lifecycle

A map of the questions a governed action raises, rather than private internal modules.

Explore the concepts →