CREATOR POSITION / OBJECT LAYER · 2026
The Artifact Layer
The industry is already building the intelligence stack. SuperArtifact.si asks a different question: what should persist at the object layer when models, agents, tools, providers and bodies keep changing?
This site does not claim to have predicted agents, tool use, MCP-style routing, operating-system integration, AI-generated apps, or robotics. Those are active engineering programs. The proposal here is narrower and more creator-oriented: treat the persistent artifact as a design unit in its own right.
Models change. Agents change. Providers change. The artifact can remain.
1. The stack is real
By 2026, major AI companies are already exposing substantial parts of the stack that once sounded speculative. OpenAI’s Agents API gives agents durable environments where they can execute code, edit files, connect to MCP servers, and produce artifacts. Google describes Android 17 as the start of a transition from an operating system to an “intelligence system”, with AppFunctions exposing app capabilities as orchestratable tools. Google DeepMind is mapping model intelligence into whole-body robot control, while NVIDIA describes an expanding full-stack ecosystem for physical AI.
At the creator layer, the word artifact is already mainstream product language. Anthropic’s Artifacts are standalone documents, code, websites, interactive components, and AI-powered apps that can be shared and customized. Anthropic reported in 2025 that users had created more than half a billion artifacts.
| Active industry layer | Example | What SuperArtifact.si does not claim |
|---|---|---|
| Agent harness + tools | OpenAI Agents API / MCP | That agents using software tools are a new observation here |
| OS capability routing | Android AppFunctions / Android MCP | That apps becoming callable capabilities were uniquely predicted here |
| Creator artifacts | Claude Artifacts | That AI-generated artifacts or remixable apps are a new category invented here |
| Physical intelligence | Gemini Robotics / NVIDIA physical AI | That connecting models to robots and machines is an unexplored endpoint |
2. The object-layer question
Once those layers exist, another design question becomes visible. A model can be replaced. An agent can be restarted. A tool server can move. A provider can disappear. A robot body can be upgraded. What, from the user’s or creator’s point of view, should remain stable enough to carry identity, capability and history across those substitutions?
SuperArtifact.si calls this focus the Artifact Layer:
Artifact Layer: the proposed conceptual layer of persistent, addressable objects through which intelligence becomes usable, attributable, revisable, composable, permissioned and inheritable across changes in underlying models, agents, tools, providers and embodiments.
This is a proposed framing, not an established industry standard. It shifts attention from the intelligence provider to the object through which capability is encountered.
3. Why artifact is a creator word
An artifact is not merely an endpoint in a software diagram. It is also something made. A drawing, file, game, instrument, page, object, tool, world, or machine can carry choices made by a creator and can persist after the moment of creation.
That is why this site stays deliberately on the artifact side rather than becoming a generic “future of agents” site. The engineering stack asks how intelligence can act. The creator question asks what kind of things we make once intelligence can act through them.
Anthropic’s product use of Artifacts is useful prior context precisely because it demonstrates that creator-facing AI output already wants to become a reusable object rather than remain trapped in chat. The SuperArtifact proposal extends that direction architecturally: an artifact can become a stable surface into capability that is not all locally contained in the artifact itself.
4. Artifact versus SuperArtifact
The site keeps the canonical 2026 definition narrow:
A superartifact is an addressable artifact that acts as a stable interface to a heterogeneous capability graph extending beyond its local representation, with at least some of that capability operationally invokable, composable, or reusable.
The distinction is therefore not “AI output versus ordinary output.” A generated image can remain a normal artifact. A hand-made object can become a superartifact if it functions as an addressable surface into wider operational capability.
| Object | Primary property | Example role |
|---|---|---|
| Artifact | Persistent made thing | Stores form, meaning, code, media, rules or memory |
| Agent | Acting process | Plans, calls tools, transforms state, completes tasks |
| Superartifact | Persistent capability surface | Stays addressable while routing to a changing graph of agents, models, tools, state and people |
5. What should survive underneath change?
If the artifact is the continuity object, raw capability is not enough. A useful artifact layer should make several properties explicit rather than burying them inside one provider’s runtime.
Stable identity
The object remains referable even when its implementation, model, tool chain, or physical carrier changes.
Where it came from
Creators, sources, transformations, dependencies, and inherited artifacts remain recoverable.
What it may invoke
Capability reach should be bounded by explicit authority rather than assumed merely because intelligence can technically reach something.
How to leave
People should be able to inspect dependencies, replace providers, export state, branch, remove capability, or return to a simpler form.
What branches survived
Remixes and mutations should not erase the routes by which an object became what it is.
How difference persists
A capability-rich environment should preserve alternative styles, methods, subcultures, and generative branches rather than compress everything toward one attractor.
6. The creator thesis
This produces a different question from the normal frontier-model race:
Can intelligence act everywhere without making everywhere the same?
The Cultural Superartifact extension answers by treating generative variation itself as something worth preserving. Provenance should not only tell us which output survived. It should preserve enough of the routes, methods, dissenting branches, tools, and transformations for future humans and machines to continue making differently.
That is the creator-oriented edge of SuperArtifact: capability matters, but so do authorship, remix, lineage, exit, style and pluralism.
7. The Marcie bridge
Dime Tower supplies the site’s infrastructure myth. The artifacts are already present. Marcie can interpret them. The route can continue when Marcie cannot. In 2026, models increasingly interpret inherited software, agents operate through software tools, and physical-AI systems extend action into machines.
The point is not that industry has failed to imagine the final endpoint. It is that carrier independence changes what can count as the durable object.
The jump is not the robot leaving the tower. It is capability ceasing to require one carrier.
From that perspective, a robot is one possible carrier. So is a phone, room, vehicle, game, document, instrument, machine, website, or future object category. The SuperArtifact question is what happens when those objects remain meaningful and addressable while the intelligence behind them keeps changing.
8. Position
SuperArtifact.si therefore does not try to out-roadmap OpenAI, Google, Anthropic, NVIDIA, xAI, robotics companies, or the rest of the AI industry.
Its narrower position is:
SuperArtifact → asks what kind of persistent objects mediate that capability
Cultural SuperArtifact → asks how those objects can preserve the ability to remain different
The engineering question is whether intelligence can act through the world. The creator question is what kind of world we choose to leave behind when it can.
Suggested web citation: Eissens, Raynor. “The Artifact Layer of the Intelligence Stack.” SuperArtifact.si, 27 September 2026. https://superartifact.si/artifact-layer/
Status: conceptual 2026 proposal. “Artifact Layer” is not presented as an established technical standard or as a claim to have originated the general concept of artifacts in AI systems.
Sources and evidence boundary
- OpenAI: Agents API overview. Source for durable cloud-agent environments, code execution, file editing, MCP connectivity, and artifact production.
- Google: Android 17 is here, 16 June 2026. Source for Android’s “intelligence system” framing and AppFunctions as orchestratable tools discoverable by agents.
- Anthropic: Turn ideas into interactive AI-powered apps, 25 June 2025. Source for AI-powered Artifacts and Anthropic’s statement that users had created more than half a billion artifacts.
- Anthropic Help Center: What are artifacts?. Source for current creator-facing Artifact behavior and examples.
- Google DeepMind: Gemini Robotics 2, 30 July 2026. Source for whole-body robot control, embodied reasoning, and adaptation across robot embodiments.
- NVIDIA: Physical AI ecosystem, 16 March 2026. Source for NVIDIA’s physical-AI models, simulation frameworks, hardware, and deployment ecosystem.
The industry sources establish that agents, callable software capabilities, creator artifacts, and physical-AI systems are active engineering categories. They do not establish the Artifact Layer or SuperArtifact framework. The framework and its normative design properties are Eissens’ synthesis and proposal.