AI Universe · Concept and physical

Concept and physical

A concept is the job and its rules. A physical thing is a product doing that job somewhere. Sorting is a concept; writing it in Java or Python is physical. Evaluation is a concept; doing it in LangSmith or DeepEval is physical.

Evaluationthe concept: inline, online andoffline checks against a barOpen sourceDeepEval · promptfooAzureFoundry evaluationAWSBedrock evaluationsGCPVertex Gen AI evalVendorsLangSmith · Braintrustlearn the concept once · the products are where it happens to live
The lavender box is the concept. The blue boxes are physical: products that do that job on one platform.

The same idea, for a whole system

The agent harness is a concept too: a loop, context, tools, state, checks, durable waits and tracing around a model call. It can be built on any platform. We built it twice, the same concerns both times; only the bindings changed.

Agent harnessthe concept: loop, context, tools, state,checks, durable waits, tracingOpen stackPython + Pydantic AIAzureContainer Apps + APIMAWSBedrock AgentCoreGCPADK + Agent Enginethe same concerns everywhere; only the bindings change
Solid blue: we have built it (see the projects). Dashed: same concept, not built by us yet.

Why this matters when you learn

When you learn……you can
the concept (what the job is, what can go wrong)judge any product that claims to do it, in an hour
only a productuse that one product, until it changes or you move platform
the concept, then one product properlyswitch to the next product by mapping, not relearning

Five kinds of physical thing

KindWhat it isExamples
Implementationdoes one concept's jobLangfuse does tracing; vLLM serves models; pgvector stores embeddings
Bundle / platformpackages many concepts under one nameMicrosoft Foundry, Bedrock AgentCore, Databricks
Protocol / standardan agreed way for two concepts to talkMCP (agent ↔ tools), A2A (agent ↔ agent), OpenTelemetry (anything ↔ tracing)
Wrappera convenience layer over another implementationLiteLLM over provider APIs; a chat app over a model
Modelthe trained weights themselvesClaude, GPT, Gemini, Llama, Qwen, gpt-oss
Patterna way of combining concepts, not a productRAG, ReAct, human-in-the-loop

Next: place the new thing →

Compute & infrastructureDataModelModel accessThe AI application (agent harness)ExperienceAssurance & operations