Foundations
6 minWhat is harness engineering (and why a GPT with tools is not enough)
Dani installed five MCPs and the agent shined in the demo — but broke on the third real task. The problem was not the model: it was the harness.
Read guide →Our school
Learn how AI agents work — from anatomy to step-by-step guides.
Anatomy of an agent
A harness combines loop, MCPs, skills, memory, and verification — click each piece to explore.
-
Agent loop
The harness core: the agent decides what to do, uses tools, and evaluates outcomes.
What you find here
Guide on harness engineering and the loop components.
Explore -
MCP servers
Model Context Protocol servers expose tools and data (DBs, APIs, files).
What you find here
Search by intent and copy ready-made config for Claude Code / Cursor.
-
Agent skills
Packaged procedures in SKILL.md — reusable expertise for specific tasks.
What you find here
Search by intent and copy the ready-made install command for Claude Code / Cursor.
Explore -
Memory & context
Context window, RAG, checkpoints, and durable state.
What you find here
Guide on choosing data and memory MCPs without over-installing.
Explore -
Verification
Tests, guardrails, permissions, and human-in-the-loop.
What you find here
Guide on permissions, verification, and a live catalog in production.
Explore -
Tools
Each MCP exposes tools; the agent picks them based on the task.
What you find here
Every card shows transport, language, and maintenance signals.
Build agents that don't break in production
Harness engineering is the software layer around the LLM: loop, MCPs, memory, and verification. These guides take you from demo to deploy — step by step, with the MCP directory as a supporting tool.
Why harness engineering?
A GPT with tools is not an agent. It is an engine without a car. Harness engineering is everything else: the loop that picks the next step, the tool catalog exposed to the model, memory across turns, and guardrails before a destructive action runs.
Most teams install five MCPs, nail an impressive demo, and break on the third real task. The problem is almost never the model — it is the harness. In guide 1 we explain the pieces and why they matter before you pick servers.
Once you know which pieces you need, the MCP directory lets you search by intent — for example an agent that reads Linear, queries Postgres, and alerts on Slack — and copy configs ready for Claude Code or Cursor.
Learning path
Four guides in sequential order. Each one assumes the previous — start with guide 1 if this is your first time with agents.
Tool design
7 minHow to choose MCP servers for your agent (without installing 40 repos)
A README with 200 MCPs paralyzes you. Here are concrete criteria to filter by intent, maintenance, and scope — before touching package.json.
Read guide →Reference implementations
6 minYour first MCP stack: Claude Code and Cursor step by step
Three MCPs — data, action, notification — and you have a minimum viable agent. Here is the example stack, the stack builder, and how to copy the config.
Read guide →Security & verification
7 minFrom prototype to production: permissions, verification, and a living catalog
The demo worked. In production the agent deleted something, burned tokens, or called a deprecated API. Here are least-privilege permissions, human-in-the-loop, and signs of an abandoned MCP.
Read guide →