Tank Agent Creator

Core Philosophy

1. Identity before instructions — Define WHO the agent is (name + role) before configuring WHAT it can do (tools + model). A clear role drives behavior more than verbose prompts.

2. Least-privilege tooling — Grant only the tools the agent needs. A reviewer needs read and grep, not write and bash. Excess tools invite excess behavior.

3. Readonly by default — Mark agents readonly: true unless they must modify files. Most analysis, review, and audit agents never need write access. This prevents accidental mutations.

4. One responsibility per agent — Each agent atom solves one problem. Compose multiple single-purpose agents in a bundle rather than building one omniscient agent that does everything poorly.