Harness Engineering: as 5 camadas do agent - memory, context, skills, agents e tools
Imagina um piloto de Fórmula 1 sentado num banco de praça. Ele sabe pilotar. Mas sem o carro, sem o volante, sem a experiência, sem o box ele não corre. O modelo de IA é o piloto. O harness é o carro inteiro. Cursor, Kiro, Claude Code, Windsurf, Cline. Essas não são "IDEs com chat". São harnesses — ambientes completos onde o modelo opera em loop, lê arquivos, executa comandos, e decide o próximo…
1. O que é harness (e o que não é)
Conceive a structure of an F1 race track. It allows drivers to drive, but without the track, the car, the experience, and the pit, the car cannot perform. The AI model is the driver. The harness is the entire track. Cursor, Kiro, Claude Code, Windsurf, Cline are not IDEs with a chat. They are harnesses - complete environments where the model operates in a loop, reads files, executes commands, and decides the next step.
ChatGPT web chat? Not a harness. It's a text box. The model answers and ends. The difference is: in a generative AI chat, you ask and it speaks. In a harness, you request and it acts. This action has five layers. Each one is the foundation for the next. If you skip a layer, the one above cannot stand. What you take away: what a harness is (and what it isn't), the maturity ladder (memory → context → skills → agents → tools), when MCP makes sense, and why RAG before organized vault is swapping a problem for two.
2. IDEs agentic: where the harness lives
Tool What is a harness? ChatGPT web Chat with model ❌ No - text window, no loop Cursor IDE with agent integrated ✅ Yes - reads code, executes, decides Kiro IDE agentic (specs, steerings) ✅ Yes - structured sessions Claude Code Terminal agentic ✅ Yes - executes in shell Windsurf IDE with Cascade ✅ Yes - multi-step flow Cline VS Code extension ✅ Yes - agent in editor Codex CLI Agent in terminal ✅ Yes - loop with tools The standard: if the model only responds with text, it's not a harness.
If it executes, sees the result, and continues, it's a harness. Inside the harness, the model operates in a cycle: 1. Reads context (files, memory, instruction) 2. Decides an action (read file, run test, edit code) 3. Executes the action 4. Sees the result 5. Decides next action or answer to you. This cycle distinguishes chat with AI from working with AI.
3. The model is the engine, not the car
Picture a car. The model is the engine. A powerful engine in a car without brakes is a problem, not a solution. The harness is the entire car: gearbox, brakes, dashboard, GPS. The engine processes. The car decides where to go. A common confusion is swapping models, thinking it solves the context problem. A better model with a weak vault still invents. A cheaper model with clean context delivers results. The model processes the context. Who defines the context is the harness.
4. The 5 layers in the right order
Order matters because each layer is the foundation for the next: Memory - the index of the library (vault hub-first, ~200 lines) Context - the workspace (only what enters the prompt now) Skills - the employee manual (domain recipes) Agents - who decides the next step (loop with policy, max 3) Tools - the universal key (access to external world) Without organized memory, agents degrade.
Without skills, code review becomes a generic KISS/YAGNI manual, not your team's. Without agents, you become a manual orchestrator. Copy/pasting between chats. Deciding each step. Without tools, you only read and write text. No Jira, no Sentry, no PR. The bottom line: without all five layers, the model works for you, not with you. Each missing layer is a task returned to you.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
