AI Team Starter Kit v2

AI Team Starter Kit v2

v2 is a ground-up redesign of the kit's operating model. v1 modeled the workspace as a human org chart: an orchestrator forbidden from doing any work, a mandatory two-agent hiring pipeline, always-binding SOPs, and hand-mirrored config files. It worked — but it paid a permanent tax in tokens, latency, and upkeep, and it grew by adding process instead of capability.

v2 rebuilds everything around what actually differentiates AI agents: tool scope, context isolation, and parallelism.

Highlights

• NEXUS works and routes. Small bounded tasks happen directly in the conversation. Delegation is reserved for three concrete triggers: parallel fan-out, context isolation, and guarded tool scopes. An orchestrator that refuses to summarize a document without spawning a subagent isn't disciplined — it's expensive.
• Function-based agents. LENS (research), QUILL (writing), and GAUGE (review — deliberately read-only) replace persona files. Each agent is a mission, a scope, testable rules, and an output contract, with real routing frontmatter that works out of the box on Claude Code and degrades gracefully everywhere else.
• Growth is one file. The hiring pipeline is gone. The scout skill walks through: can an existing agent absorb this → is it a skill or an agent → optionally research the domain with LENS → write the file from the template → smoke-test, review, one changelog line. The changelog is the roster.
• Skills over SOPs. Procedures load on demand when a task matches. The only standing context is a small policies/ folder — because every always-on rule costs attention forever.
• AGENTS.md is generated. CLAUDE.md is canonical; sync scripts (sh + ps1) render AGENTS.md from it, so the two can never drift. Optional validate scripts check workspace hygiene, frontmatter, and freshness.

Migrating from v1

There is no in-place migration — v2 is a different operating model. Start a fresh copy of the kit, port your owner.md profile, and convert any SOPs you actually use into skills (the write-procedure skill shows how). v1 remains available at the v1 tag.

What carried over

The names (NEXUS, SCOUT, LENS), the owner profile, the token economy, and the failure-handling rule: report failures with a diagnosis — never silently retry.