Claude Code Agent Teams: skill factory with 7 agents

What this system is

claude-agent-teams-skill-factory is a hybrid architecture combining Agent Teams and subagents in Claude Code to build a factory of reusable instructions (skills). With a single prompt, the team researches, writes, debates, compresses, validates and publishes — without manual intervention.

The result: SKILL.md skills published to the Hermit portal and available locally for Claude, Cursor and any compatible agent.

The 7 agents

Agent Team (4 Claude Code sessions):

Agent	Type	Responsibility
Lead	Team lead	Orchestrates the full pipeline, launches subagents
Architect	Teammate	Writes SKILL.md as instructions for Claude
Reviewer	Teammate	Checks if Claude would execute the instructions correctly
Optimizer	Teammate	Removes everything Claude already knows

Lead’s Subagents (no own session):

Agent	Type	Responsibility
Investigator	Subagent	Fetches official docs, returns research
Validator	Subagent	Runs validate.sh, returns PASS/FAIL
Publisher	Subagent	Publishes to the portal and verifies

Why two layers

The distinction is not cosmetic — it defines available tools, assigned models and cost:

• Teammates communicate with each other via Mailbox. They need to debate to build the artifact well. Agent Teams assigns Opus to them automatically.
• Subagents don’t have access to the Task Tool, can’t launch other agents, and inherit the Lead’s model (Sonnet). They execute a scoped task and return a result.

Design rule: do the agents need to talk to each other to do the job well? Yes → Agent Team. No → Subagents.

The model of models

Caution

Agent Teams automatically assigns Opus to all teammates regardless of what you configure in CLAUDE.md or the spawn prompt. The Lead inherits the model from your active session. Subagents inherit the Lead’s model. Haiku was never used in any demo — confirmed data from 37 analyzed calls.

Component	Actual model
Lead	Sonnet (inherits from your session)
Architect, Reviewer, Optimizer	Opus (always, automatic)
Investigator, Validator, Publisher	Sonnet (inherits from Lead)

The only possible saving: launch Investigator, Validator and Publisher without team_name to keep them as subagents on Sonnet.

What the system produces

A complete run produces a validated and published SKILL.md:

skills/skill-name/
└── SKILL.md    ← instructions for Claude, < 500 lines

The skill becomes available in two places:

• Hermit portal: http://localhost:8080
• Local: ~/.claude/skills/skill-name/SKILL.md

Claude detects portal skills automatically and picks the right one without the user specifying which skill to use.

Technologies

Component	Technology
Agent framework	Claude Code Agent Teams (experimental)
Skills portal	Hermit (self-hosted, Docker)
Skill validation	validate.sh + skills-ref (optional)
Publishing	publish.sh + Hermit API
Teammate model	Claude Opus 4.6 (automatic)
Lead + subagent model	Claude Sonnet 4.6 (inherited)

→ Quick start: installation and configuration