AI Products Are Splitting Into Three Layers: Belief, Composition, and Deployment

Three releases attracting developer attention this week sit in different categories. Ataraxos is a reinforcement-learning system for Stratego. FLUX 3 Image is a controllable generation and editing model. ChatGPT Sites is a way to create and publish interactive websites and lightweight apps.

They are not integrations and the organizations do not present them as one stack. Yet they reveal the same product direction: useful AI systems are exposing intermediate state instead of asking users to trust a single opaque output. Ataraxos represents uncertainty with sampled hidden boards. FLUX 3 represents composition with boxes and semantic identifiers. Sites represents deployment with versions, access controls, connected-app permissions, and publish settings.

The emerging stack has three layers: belief, composition, and deployment. The model’s output still matters, but the interfaces around it increasingly determine whether a result can be inspected, corrected, and safely used.

Layer one: belief state

Ataraxos cannot see the identities of most opposing Stratego pieces. Its belief network predicts a distribution over the hidden board from observable play. Test-time search then evaluates moves across samples from that distribution.

This is a useful abstraction beyond games. An agent rarely has a complete world state. It has documents that may be stale, users who omit context, tools that fail, and permissions that reveal only part of a system. Treating the current prompt as complete truth creates brittle plans.

A product-grade agent should be able to say what it knows, what it infers, and which alternatives remain plausible. Those states do not need to look like a Stratego board. They may be confidence-weighted hypotheses, missing-field flags, contradictory records, or simulated outcomes. The architectural point is to keep uncertainty explicit enough that later actions can account for it.

The limitation is calibration. A beautifully structured belief state can still exclude the truth. Products need tests for how quickly a system updates after contradictory evidence and whether consequential actions require stronger evidence than reversible ones.

Layer two: composition state

FLUX 3 Image’s bounding-box interface turns a visual request into a set of persistent elements with coordinates and descriptions. A person or agent can move one item, replace another, and attempt to preserve the rest.

This is the media equivalent of structured application state. Prompt-only generation is easy to start but difficult to revise. When every change requires a new paragraph, the desired design exists only implicitly in language and in the user’s memory. Boxes and identifiers provide handles.

The same principle applies to documents, video, code, and data products. A reliable system should expose the objects it is manipulating: scenes, components, claims, sources, tests, and dependencies. Users should be able to revise one part without renegotiating the entire artifact.

The interface also creates a role for agents that is more concrete than “be creative.” An agent can plan a layout, detect collisions, compare variants, and preserve invariants. That work is testable. It is still possible for the renderer to violate the plan, but the plan itself is visible.

Layer three: deployment state

ChatGPT Sites adds the state that begins after generation. OpenAI’s help documentation describes private previews, saved versions, production URLs, sharing options, collaborator roles, custom domains where available, connected apps, and scheduled updates. It says Sites is in public beta for workspaces, Plus, and Pro accounts, with availability still rolling out and plan-specific limits.

For Business and Enterprise workspaces, a Site can request read-only access to a visitor’s own connected apps when administrators and the visitor authorize it. Each visitor uses their own accounts and permissions. Public publishing and connected data are therefore different control planes.

This matters because a generated interface is not a product until its audience, data access, update path, and failure behavior are defined. A preview can be correct for the creator and wrong for a teammate whose permissions expose different records. A scheduled task cannot simply inherit a visitor’s interactive connection. A public link can reveal information that was safe inside a workspace.

The Hacker News submission for Sites showed 239 points and 53 comments when checked on October 3. That attention reflects interest, not a security audit. The official documentation itself repeatedly instructs users to review access, sensitive data, forms, links, and visitor behavior before publishing.

Why these layers are converging

One-shot generation was sufficient when AI outputs were disposable suggestions. As systems take actions and produce durable artifacts, three questions become unavoidable:

1. What assumptions led to this decision? That is belief state.
2. What parts make up this artifact, and what may change? That is composition state.
3. Who can use it, with which data, and how is it updated? That is deployment state.

These questions create a new competitive surface. Model quality remains important, but a slightly weaker model with strong state, revision, and permission interfaces may outperform a stronger model embedded in an opaque workflow.

The layers also support different tests. Belief can be evaluated for calibration and robustness. Composition can be evaluated for constraint compliance and edit isolation. Deployment can be evaluated for access control, reproducibility, rollback, and observability.

Practical architecture for builders

Start by separating the three states in storage. Do not bury all of them in a conversation transcript. Record evidence and hypotheses for decisions; structured objects and invariants for the artifact; versions, principals, permissions, and environment information for deployment.

Make transitions reviewable. A person should be able to inspect the beliefs used to approve a composition, then inspect the exact version that moved from preview to production. If an agent runs the transition, keep the same record and define which steps require confirmation.

Use different retry rules. Recomputing a belief from read-only evidence is usually reversible. Rendering a new draft is often low risk. Publishing, sending, charging, or writing through a connected service is not. Treating all three layers as “another model call” erases the distinctions that make automation safe.

Finally, test the seams. A well-calibrated belief model can feed an impossible layout. A perfect composition can expose confidential information when deployed. A secure deployment can faithfully publish an unsupported claim. End-to-end evaluation must include handoffs, not only component scores.

Limitations

This three-layer framework is an editorial synthesis, not a standard proposed by Nature, Black Forest Labs, or OpenAI. The products differ in purpose, maturity, and evidence. Ataraxos is a research system in a simulated game. FLUX 3’s launch page is vendor documentation. Sites is a public beta product with changing limits and controls.

Explicit state also has costs. It requires storage, schemas, migration, privacy policy, and user-interface design. Too much visible structure can overwhelm users or encourage false precision. A belief score is not truth, a bounding box is not a guarantee, and a permission toggle is not a complete security model.

The direction is still meaningful. AI products are moving from impressive outputs toward inspectable processes. The teams that make uncertainty, composition, and deployment legible will be better positioned to build systems that people can correct—and trust for more than a demo.