The AI Taxonomy.
Terms  /  Architecture  /  Model-Centric AI
05 · Architecture

Model-Centric AI

An architecture where the model is expected to infer state, intent, plan, and memory from raw context.

Model-Centric AI is the default approach in most AI applications today: put everything in the prompt and let the model figure it out. The model is expected to read the conversation history and infer the current plan. It's expected to scan previous messages and remember what was decided. It's expected to interpret ambiguous instructions and determine intent. This works for simple tasks, but it breaks down as complexity grows because every responsibility piled on the model increases Reasoning Load, Context Debt, and token cost. The model becomes the bottleneck for everything.

Example
A typical chatbot-based coding assistant is Model-Centric. The user types instructions, the system appends them to the full conversation history, and the model receives everything. At turn 50, the model is processing 40,000 tokens to determine: what's the current task? What was already tried? What conventions should be followed? What files exist? All of this is inferred from raw context rather than provided by a system layer. The model works harder, costs more, and becomes less reliable with every turn — because it's doing the system's job in addition to its own.

More in Architecture

System-Centric AI

An architecture where planning, state, memory, and control live outside the model.

Tool Isolation

Memory and context don't cross tool boundaries — each tool is a silo.

The Expertise Trap

Context and knowledge locked on one person's machine, inaccessible to the team.

Agent Fragmentation

Agents operating in disconnected silos with no shared state, duplicating work and contradicting each other.

Unified Context

A single, shared knowledge layer accessible to all agents and tools. The cure for Tool Isolation.

Context Portability

The ability to move knowledge between tools, agents, sessions, and team members without loss. The cure for The Expertise Trap.

Stateful, Inspectable Memory

A collaborative memory space where multiple agents and humans contribute to and draw from the same knowledge base — with full history and traceability.

Context Handoff

The ability to transfer full working context from one agent, session, or tool to another without loss. A specific operation enabled by Context Portability.