Containment:
When it goes wrong:
Human veto:
AI systems by artifact left behind and execution control A scatter chart. The horizontal axis is the artifact left behind, running through six waypoints: Chat Log, Versioned Artifacts, Media Library, Change Specific, Change Anything and Become. The vertical axis runs from more deterministic at the bottom to more autonomous at the top. Systems trend diagonally from lower left to upper right. The lower right is empty, because an undeclared or absent artifact cannot be developer-sequenced. Persistence rises from Chat Log through Change Specific. At Become the output may still persist, but the software that produced it does not exist as reviewable code and may differ on each use. AI Systems by Artifact Left Behind and Execution Control Placement is editorial, not measured Chat Log A transcript, and whatever you copy out Versioned Artifacts Buildable and revertable; read before they count Media Library Durable, but an island by default; little reads it yet Change Specific Records other systems depend on, like a database or CRM Change Anything Traces wherever it reached, using any tool or API Become Output may persist; the thing that made it does not — and may differ next time you open it Artifact left behind More autonomous Model chooses actions and sequence More deterministic Developer defines actions and sequence Execution control Deep research modes Claude.ai, ChatGPT web chat IDE chat pane (suggest only) Prompt templates CI linters Dependabot, Renovate Cursor Tab GitHub Copilot CodeRabbit Codex cloud agent Claude Code, Codex CLI Open- Hands Devin Sora, Veo, Midjourney Batch image generation MCP / tool-calling apps — lands wherever they are wired Data enrichment Zapier “When X happens, update system Y” n8n workflows Support agents IT agents CRM agents OpenHands Voice agents (booking, intake) Trading and pricing bots Finance and procurement agents (ERP) OpenClaw AI-native applications “saved” is itself generated text Nothing lands here: an undeclared or absent artifact cannot be developer-sequenced AUGMENTED model fills a slot in someone else's sequence HARNESSED model plans, inside a declared rig with a review gate UNHARNESSED model plans; the rig is undeclared FAST DETECT / FAST REPAIR wrong output is visible at once, and cheap to undo FAST DETECT / SLOW REPAIR you notice quickly; unwinding it takes days SLOW DETECT / FAST REPAIR nobody looks for a while, but a fix is trivial SLOW DETECT / SLOW REPAIR plausible, unnoticed, and load-bearing by the time it surfaces VETO BEFORE IT COUNTS rejection is free: nothing has happened yet — including chat, where you simply do not use it VETO AFTER THE FACT it already counted; undoing it means a compensating action, if anyone looks NO VETO you cannot intervene in time, and cannot inspect what acted