# Agent Workflow Canvas > An interactive enterprise workflow prototype by Kevin Astuhuaman. It demonstrates how non-technical operators can configure, inspect, approve, and recover AI-assisted work without surrendering control. ## What this proves - Enterprise AI interface design: a dense but legible builder, inspector, and execution trace. - Human-in-the-loop product judgment: approval is a durable run state tied to evidence and committed inputs. - Agent reliability: deterministic validation, confidence routing, fail-closed permissions, and explicit recovery. - Technical fluency: Astro, TypeScript, browser tests, structured metadata, and an auditable public implementation. ## Inspectable scenarios 1. Invoice mismatch: the workflow pauses before the ERP commit and requests accountable finance approval. 2. Clean invoice: the workflow follows an automatic branch and creates a draft without releasing funds. 3. Permission revoked: the connector fails closed, preserves the uncommitted draft, and recovers with scoped access. ## Product decisions - The canvas and execution trace share selection context. - Consequential actions sit behind explicit policy boundaries. - AI extraction returns evidence for every field. - Synthetic, deterministic scenarios make every important state reviewable. - The public repository contains no employer code, customer data, production credentials, or proprietary prompts. ## Links - Live product: https://kevinastuhuaman.github.io/agent-workflow-canvas/ - Source: https://github.com/kevinastuhuaman/agent-workflow-canvas - Structured project record: https://kevinastuhuaman.github.io/agent-workflow-canvas/project.json - Decision 001: https://github.com/kevinastuhuaman/agent-workflow-canvas/blob/main/DECISIONS/001-approval-is-a-state.md - Decision 002: https://github.com/kevinastuhuaman/agent-workflow-canvas/blob/main/DECISIONS/002-trace-and-canvas-share-context.md - Builder Stack: https://kevinastuhuaman.github.io/ai-product-builder-stack/ - Portfolio: https://portfolio.kevinastuhuaman.com