What is an AI Harness
A harness is the system around the model: persistent context, memory, and explicit rules. Claude Code, Cursor, Windsurf, and OpenCode are harnesses; Claude, GPT, and Gemini are models.
Felipe Luis Salgueiro / How to build your own AI harness
How to build your own AI harness with Claude Code
Lesson 01 of the Building with AI series. A harness turns a model into a system with context, memory, rules, and repeatable execution.
Lesson 01 · June 4, 2026 · 120 min · ~150 professionals live · NoCode Startup
An introduction to the harness as the system around the model: context, memory, rules, and repeatable execution.
Why this topic now
- Using AI at work is already a commodity; building with APIs, no-code, and isolated automations is the next stage.
- This lesson addresses a third stage: operating AI with context, memory, and specialization rather than relying on loose prompts.
Format
Monthly live lesson at NoCode Startup. Recording is exclusive to students; this page records public concepts and materials in text for reference.
Explore NoCode StartupWhat the student learns
- How Claude Code differs from chat and AI IDEs.
- Harnesses and their three foundational layers.
- CLAUDE.md, squads, orchestrators, workers, hooks, subagents, and MCP.
- Prompts inside agent systems, multi-model work, and permissions.
What they can do
- Install and configure Claude Code.
- Create a structured CLAUDE.md, a simple squad, and an invokable skill.
- Recognize harness architecture and leave with an expandable skeleton.
Concepts covered in detail
The three layers of a Harness
Context defines identity and limits; memory records decisions and progress; execution turns knowledge into repeatable skills and workers.
CLAUDE.md — the agent manual
This file holds identity, scope, conventions, limits, and operations requiring confirmation. Cascading reading enables global, project, and area-specific rules.
Squads — specialized agents
Squads have their own scope, context, memory, and skills by area. Specialization reduces irrelevant context and makes execution more predictable.
Orchestrators and Workers
The orchestrator decides and delegates; workers execute specific tasks and report. The separation avoids loading the whole system for every task.
Hooks — declarative automation
Hooks run automatically at defined points in a flow. When a rule is ‘whenever X happens, do Y’, the harness must execute it rather than merely remember it.
Skills — encapsulated behaviours
A skill documents a reusable behaviour with objective, steps, inputs, outputs, and guardrails. Small, versioned skills preserve predictability.
MCP — Model Context Protocol
MCP connects external tools through an open protocol. The lesson also covers context hygiene: keep only the integrations needed for a task.
Multi-model — the right AI for each task
Models have different strengths. Choose by task and criterion, and keep the ability to swap models without rewriting the system.
Permissions — what an agent can and cannot do
Automatic capabilities, actions requiring confirmation, and forbidden actions must be explicit. Destructive operations and production require human approval.
Lesson structure
- Opening: see the destination before the path.
- Under the hood: architecture of a real harness.
- Live building: first context, squad, and skill.
- Questions and next paths.