Build your own harness
On its own, an AI model only writes text. What turns it into an agent that can finish a long piece of work is everything around it: a harness. It is the work loop that lets the model plan, use tools, read and write files, look up your company’s guides and hand out tasks, step by step, until it delivers the result.
The best-known coding assistants are exactly that: a good model inside a good harness. In NappAI the harness is the Deep Agent component, and you decide what to equip it with:
| Piece | What it gives the agent | How it is connected |
|---|---|---|
| Model | The intelligence that decides each step. | A model component in Model. |
| Tools | Concrete actions: search the internet, read a spreadsheet, send an email, query a database. | Any component with a Tool output in Tools. |
| MCP servers | The tools other applications publish with the MCP standard, ready to use. | A server from the MCPs category, or MCP Connection, in Tools. |
| Skills | Your company’s know-how: how a task is done, in what order and in what format it is delivered. | Stored Skill or a NappAI skill in Skills. |
| Plan and files | A task list the agent ticks off and working files where it keeps notes. | Built in. |
| Subagents and code environment | Specialised helpers to delegate to, and an isolated computer to run code on. | Optional, in the advanced options. |
The result is an agent made to measure for one job: with just the tools it needs and your company’s method written as skills, without programming.
This video builds a complete harness: an assistant that prepares the naming study for a new business.
What you see in the video
- In My Skills, the
naming-studyskill is created: the agency’s method for finding names, in four steps and with the format of the report. - In the flow, a Deep Agent with a DeepSeek model gets short instructions: before starting, check whether a skill explains the task and follow it.
- Its first tool is Web Search (SearXng), to search the internet with no keys.
- The second one is an MCP server, Find-A-Domain, from the MCPs category: it checks whether a domain is free. Its
check_domaintool is chosen and its data is marked for the AI to fill in. - Turning on Enable Skills shows the Skills connection, and a Stored Skill with the new skill is connected to it.
- In the Playground, the agent is asked for the naming study of a specialty coffee shop in Manchester. It looks up the competitors, suggests names, checks their domains, replaces the taken ones and delivers the report in the format the skill asks for. In the video it drops the names with both domains taken, replaces them with others and recommends “Cafebrote”, with both the .com and the .net free.
The full example, with the texts to copy, is in Example: naming study.
In this section
- The pieces of the Deep Agent: what each option does and when to turn it on.
- Skills: how to write a skill the agent uses well.
- Example: naming study: the flow in the video, step by step.