The pieces of the Deep Agent
The Deep Agent comes with what a long job needs and leaves the rest as pieces you turn on in its advanced options. The rule is simple: turn on only what the job needs. Every extra piece is information the model reads at every step, and that costs time and money.
What is built in
- Work plan. Faced with a job that has several parts, the agent writes a task list and updates it as it goes. That way it does not get lost in long jobs or skip steps.
- Working files. The agent has its own space where it keeps notes, drafts and intermediate results while it works, instead of carrying everything in the conversation. That is also where it reads its skills.
- Status lines. With Tool Status on (the default), the chat shows which tool is being used at each moment. In the Playground, the Executions list shows every step with what went in and what came out.
Tools
Any component with a Tool output connects to the Tools input: web search, spreadsheets, email, databases, your storage, a Python Code with your own logic, or even another agent. A few ideas so the agent uses them well:
- Connect only the tools for the job. With twenty tools available the agent takes longer to decide and makes more mistakes than with four.
- In the Tool tab of each component, Tool Description is what the agent reads to decide when to use it. Write it for the agent: what it does and when it helps.
- Data the agent must decide is marked with the AI button next to the field (“Defined by AI”). Fields you do not mark keep the value you type.
MCP servers
MCP is an open standard that many applications use to offer their tools to any agent. In the palette, the MCPs category brings together more than a hundred ready servers; you can also connect any other one with MCP Connection. They connect to Tools just like a normal tool: you choose the server’s tool and mark the data the AI fills in. See Use MCP servers.
Skills
A skill is a work guide the agent reads when the task matches its description: the steps of a procedure, your company’s rules or the format of a report. It is turned on with Enable Skills, which shows the Skills connection. You can connect your own skills with Stored Skill, or skills NappAI has prepared, from the Skills family of the palette, for services such as Slack, HubSpot, Jira or Google Workspace (ManagedSkill).
Turning on Enable Skills also turns on Use memory, so the agent remembers what it has already read. How to write a good skill: Skills.
Subagents
With Enable Subagents the Subagents connection appears, where you connect the Agent output of other agents. The Deep Agent hands them parts of the work, and each one works in its own conversation without filling up the main one. Use them when the job has very different parts (research, writing, reviewing) or when one part produces a lot of information the main agent does not need to see in full. Give each subagent a clear Agent Name and Agent Description: that is what the main agent reads to choose.
Code environment
With Enable Sandbox the Sandbox connection appears, where you connect an isolated computer such as Nappai Sandbox. The agent can run code there to process data, make calculations or produce files and charts, without touching anything on your system. Connected skills are copied to that environment too.
Review before delivering
With Enable Rubric, before finishing, a reviewer checks the work against your criteria (one per line) and, if something is missing, the agent keeps working. It is the way to demand a specific result, for example “the report cites at least three sources”.
Rules that are never broken
Guardrails checks every request and every action before a tool is run, and blocks the ones that break your rules. It matters in a harness with tools that change things (send, delete, pay). See Guardrails.
Which piece to use
| If you need the agent to… | Use |
|---|---|
| Do things in other services | Tools or MCP servers |
| Work with your company’s method | Skills |
| Split a very large job into parts | Subagents |
| Calculate, transform data or produce files | Code environment |
| Deliver only when the work meets a minimum | Review (Rubric) |
| Never do certain things | Guardrails |
| Remember the conversation | Use memory |