Each agent is a small Markdown file with a YAML header (name, description, allowed tools) followed by plain-language instructions — no special framework beyond what Claude Code provides natively.

network-sec

Domain	Firewall, VLANs, NAT, VPN, reverse proxy
Triggered by	Any networking or reverse-proxy task
Core principle	Least privilege — never an "any/any" rule

infra-iac

Domain	Declarative provisioning, configuration automation
Triggered by	Provisioning, playbooks, automation
Core principle	Idempotence, everything versioned in Git

compliance-lab

Domain	Hardening (generic security baselines), reproducing professional scenarios
Triggered by	Hardening, compliance, remediation work
Core principle	Generalize the problem before documenting it

pm-docs

Domain	Project tracking, technical documentation
Triggered by	Task tracking, documentation writing
Core principle	It documents and tracks — it doesn't decide the work split

research

Domain	Technical research, comparisons
Triggered by	Questions needing current, sourced information
Core principle	Factual synthesis, sources cited, no unsourced opinion

How an agent gets picked

The main agent reads each agent's description and figures out which one fits the task — or I can force a specific one explicitly if I already know which specialist I want.

Adding a new agent

1. Write a new agent definition file with a clear name.
2. Give it a precise description — that's the field the automatic routing relies on.
3. List only the tools it actually needs.
4. Write its role, principles, and expected output format in plain language.
5. Test it in isolation before folding it into normal use.