How to Make an AI Agent With Claude

Claude does not have one "make an agent" button. It has several modes that support agent-like behavior, and the right choice depends on the task. Claude Projects give you persistent instructions and shared memory for a team. Claude Code runs agentic development tasks from the terminal. The Anthropic API lets you build a custom agent with defined tools and prompts. Tools like fenbs supply reusable context and the record of completed work that any of these setups can read, so each session starts from the current state rather than a blank slate.

The common thread across all three is the same: a narrow task definition, precise instructions, and a testing phase before the agent handles real work. Claude is a capable model, but an agent built on top of it is only as reliable as its instructions and tool permissions.

Choose your Claude setup

• Claude Projects (claude.ai): persistent instructions and knowledge files shared across conversations in a team. Good for a configured assistant that multiple people use. Claude Code Projects added shared memory on September 17, 2026.
• Claude Desktop with MCP: connect external tools (databases, APIs, files) using the Model Context Protocol. Good for an assistant that needs to read from or write to outside systems from a desktop interface.
• Claude Code (terminal): agentic coding and file tasks. Reads AGENTS.md for project-specific instructions. Good for development work where the agent needs to read, write, and run code.
• Anthropic API: build a fully custom agent with the model as the engine, your own tool definitions, and your own orchestration logic. Good for production systems and embedded assistants.

Write the system instructions

No matter which setup you use, the system instruction is the most important configuration. It tells Claude its role, the rules it follows, the tools it can use, and what it should do when it encounters something outside its scope. A precise system instruction produces a predictable agent.

Write the role first: "You are the research assistant for a three-person financial advisory team. You look up information, summarize documents, and draft research notes." Then add the rules: "Always cite the source. Never make financial recommendations. If the answer requires current market data, say so and note what source to check."

Add the escalation rule: what Claude should do when it cannot handle a request with the tools and information it has. An agent that confesses its limits is more useful than one that guesses confidently. AI agents for beginners covers the broader concepts if this is your first agent build.

Grant the tools the agent needs

For Claude Desktop with MCP, you configure which servers (tools) Claude can connect to: a file system server, a database, an API. Each MCP server provides a set of actions. Grant only the tools the agent needs for the defined task. Write the tool list in the project brief before configuring so the setup matches the task.

For the API, you define tools in the request schema: a function name, a description Claude uses to decide when to call it, and the parameters it accepts. Write descriptions that are specific enough for Claude to distinguish between tools. A vague description leads to the wrong tool being called. AI agent architecture covers how tools fit into a larger agent design.

Test with known inputs

Before running the agent on real work, build a set of test inputs. Include the common cases, two or three edge cases, and at least one input the agent should decline or escalate. Run each one and check the output. The most important test is the escalation case: does Claude correctly recognize when it should not proceed?

After testing, update the instructions to address anything that came up. Common issues: Claude attempts a task outside its defined scope, uses a tool when the instructions say to wait, or produces a format that does not match the expected output. Each fix is a more precise instruction, not a different model.

Keep the approved work context current

An agent set up once and never maintained drifts from the work it is meant to support. The project context — the decisions, the rules, the current tasks — needs to stay current. When a decision changes, update it. When a task is done, mark it. When a new rule is agreed, add it.

A current context record is what makes an agent useful over weeks and months rather than just in the first session. Types of AI agents covers how different agent architectures handle memory and context.

How to do this with fenbs

1. Create a free account and add a project for the work this agent supports.
2. Write the role, approved tools, and standing rules as Decisions and rules in the project.
3. Connect Claude over MCP from Integrations and ask it to read the project context before each session.
4. Add tasks for each build phase: instructions draft, tool configuration, testing, and first real run.
5. After each session, ask Claude to save where it left off. The next session starts from the current state.