AI Architecture<\/p>\n<\/div>
And what is the difference between an LLM, an agent, and an agent harness?<\/p>\n<\/div>
There is a lot of sloppy language floating around AI right now.<\/p>\n
People say \u201cagent\u201d when they mean \u201cLLM.\u201d They say \u201cagent system\u201d when they mean \u201csome code that calls GPT.\u201d And \u201cagent harness\u201d sounds like something you would buy at Home Depot.<\/p>\n
Let\u2019s clean it up.<\/p>\n<\/div><\/div><\/div><\/div><\/div>
An LLM by itself is not an agent.<\/p>\n
It is a model. You give it input. It gives you output.<\/p>\n<\/div>
The model is powerful, but by itself it is mostly a very smart text transformation engine.<\/p>\n<\/div>
That is useful. But it is not an agent yet.<\/p>\n<\/div><\/div><\/div><\/div><\/div>
An agent starts when you wrap the LLM in behavior.<\/p>\n
The key difference is that the agent has some version of a goal, a loop, a policy, and tools.<\/p>\n<\/div>
This can be very simple. It is often just Python or TypeScript, a prompt, an API call, and a loop.<\/p>\n<\/div><\/div><\/div><\/div><\/div>
Most agents are not magic. They are software.<\/p>\n<\/div>\n
| Piece<\/th>\n | What it does<\/th>\n | Commonly built with<\/th>\n<\/tr>\n<\/thead>\n | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LLM<\/td>\n | Reasoning, planning, language generation<\/td>\n | OpenAI, Anthropic, Gemini, Llama, Mistral<\/td>\n<\/tr>\n | ||||||||||||
| Agent loop<\/td>\n | Decide, act, observe, repeat<\/td>\n | Python, TypeScript, LangChain, AutoGen, CrewAI<\/td>\n<\/tr>\n | ||||||||||||
| Tools<\/td>\n | Let the agent take action<\/td>\n | APIs, databases, code execution, web search<\/td>\n<\/tr>\n | ||||||||||||
| Memory<\/td>\n | Stores context and prior information<\/td>\n | Redis, Postgres, vector databases<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n You can build this from scratch. Or you can use agent frameworks. The frameworks help with the shape of the system, but they do not remove the need to understand what is happening.<\/p>\n<\/div><\/div><\/div><\/div><\/div> 3. The Harness: The Part That Makes It Real<\/h2><\/div>Most agent demos look impressive because the happy path works.<\/p>\n Production is not the happy path.<\/p>\n The harness is the layer around the agent that handles reliability, safety, monitoring, testing, and operations.<\/p>\n<\/div> The harness answers the uncomfortable questions:<\/strong> What happens when a tool fails? Who approves sensitive actions? How do we know what the agent did? How do we test whether the new version is better? How do we prevent mistakes at scale?<\/span><\/div><\/div> This is where the work gets serious.<\/p>\n<\/div><\/div><\/div><\/div><\/div> What Lives in the Harness?<\/h2><\/div><\/i><\/div> Execution control<\/h3><\/div>\nRetries, timeouts, rate limits, and error handling.<\/div><\/div><\/div>\n <\/i><\/div> Guardrails<\/h3><\/div>\nPermissions, policy checks, output validation, and approval flows.<\/div><\/div><\/div>\n <\/i><\/div> Observability<\/h3><\/div>\nLogs, traces, metrics, cost tracking, and debugging tools.<\/div><\/div><\/div>\n <\/i><\/div> Evaluation<\/h3><\/div>\nTest sets, benchmarks, scoring, and regression checks.<\/div><\/div><\/div>\n <\/div><\/div><\/div><\/div><\/div><\/div> What the Harness Is Built With<\/h2><\/div>State and storage: Postgres, MongoDB, Redis<\/div><\/li>\n Vector memory: Pinecone, Weaviate, Chroma, pgvector<\/div><\/li>\n Logging and tracing: LangSmith, Helicone, Arize, Weights & Biases<\/div><\/li>\n Monitoring: Prometheus, Grafana, Sentry<\/div><\/li>\n Cloud infrastructure: AWS, GCP, Azure<\/div><\/li>\n Custom policies: usually your own code<\/div><\/li>\n<\/ul> So no, the harness is usually not one magical library. It is a combination of libraries, services, and custom rules that match your business.<\/p>\n<\/div><\/div><\/div><\/div><\/div> Build vs. Buy<\/h2><\/div>The practical answer is: buy what is generic, build what is specific.<\/p>\n<\/div>\n
|