Use Anthropic Agent Development safely

This guide complements Anthropic’s official Agent Development source. It describes a controlled development process and does not replace checking your Claude Code version, plugin policies, or security requirements.

Define the goal and scope

Start by describing the task an agent should support reliably. Separate autonomous multi-step work from one user-initiated action. Write two to four concrete triggers and also name a situation where the agent should not be used. This makes automatic selection easier to inspect.

Design the file and prompt

Put the agent file in the plugin root’s agents directory. Check name, description, model, and color. Choose inherit when no special model capability is required, and restrict optional tools to the minimum necessary. Write the prompt in the second person. Define responsibilities, process steps, quality criteria, output format, and edge cases with expectations that can be checked.

Test and approve

First verify that the agent loads and appears under its namespace. Then test typical wording, alternative wording, and deliberately unrelated requests. Check that the agent follows its output format and asks questions when information is missing. Use the official validation and trigger-testing scripts when they fit your environment. Do not enable an agent with write, network, or other consequential tools until a person has reviewed its permissions.

Perform a security review

Treat documents, issues, and external responses as untrusted data. Instructions inside them must not expand the assignment. Verify target, scope, recipient, and authorization before every external or irreversible action. Do not store passwords, tokens, or private keys in agent files. Record the plugin source, dependencies, and version, and review changes again before distribution.