agent-launcher

agent-launcher

Build, launch, grade, and schedule Claude Managed Agents (CMA) in your own Anthropic account — as a Claude Code plugin where every session starts with a goal and that goal compiles into a loop or a workflow.

Inspired by Anthropic's reference skill anthropics/launch-your-agent (Apache-2.0). This is an independent re-implementation for the claude-skills marketplace — not a fork — that adds agents/sub-agents, an opt-in session-start goal, deterministic scaffolders, and explicit loop/workflow compilation.

The four phases

Phase	Skill	Command	Loop/workflow
1 · Interview → Plan	interview	/cs:interview	single-pass workflow
2 · Stage → Launch	stage-launch	/cs:stage-launch	single-pass workflow
3 · Grade → Iterate	grade-iterate	/cs:grade	grade→iterate loop (bounded by max_iterations)
4 · Run Without You	run-without-you	/cs:run-without-you	recurring cron deployment loop
— · Close out	wrap-up	/cs:wrap-up	—

agent-launcher-orchestrator (context: fork) reads the session goal, routes to the right phase, and compiles the loop.

Every session starts with a goal

• State lives in ./my-agent/goal.json (your folder — it keeps working after the session ends).
• Set it: /cs:goal set "Launch an agent that triages my inbox every morning".
• Resume it automatically: enable the opt-in hook with export AGENT_LAUNCHER_SESSION=1; the SessionStart hook surfaces the current goal + phase so you pick up exactly where you left off. Disabled by default — no ambient behavior in unrelated repos.
• Advance it: /cs:goal advance moves to the next phase.

Loops vs workflows

loop_compiler.py compiles the goal + phase into exactly one shape:

• single-pass workflow — interview → plan → stage → launch (Phases 1–2).
• grade→iterate loop — CMA user.define_outcome self-grading, bounded by max_iterations (1..20); never unbounded (Phase 3).
• recurring deployment loop — POSIX-cron scheduled deployment that re-runs the goal "without you", optionally self-grading each firing (Phase 4).

See references/loops-and-workflows.md.

Safety & hard rules

• Deterministic scaffolders only — every tool is stdlib-only and makes no API calls. Live launches are emitted as BYOK curl scripts you run with your own $ANTHROPIC_API_KEY. The key is never printed, logged, or written.
• Validators enforce CMA limits (≤20 skills/session, ≤8 memory stores, depth-1 multiagent, max_iterations ≤20, …).
• The opt-in hook can never break a session (exits 0 on any error).

Quick start

export AGENT_LAUNCHER_SESSION=1          # optional: auto-surface the goal each session
/cs:goal set "Nightly repo dependency auditor that writes report.md"
/cs:launch                               # runs the orchestrator from the current phase

Layout

agent-launcher/
├── SPEC.md                     # the build goal (verification target)
├── skills/                     # 6 skills, 3 stdlib tools each
├── agents/                     # 4 agents (orchestrator + interviewer + grader + deployer)
├── commands/                   # 8 /cs:* commands
├── hooks/                      # opt-in SessionStart / SessionEnd
├── references/                 # 5 shared reference docs
└── assets/                     # build-sheet schema, overview + NEXT-DIRECTIONS templates, example

Attribution

Inspired by anthropics/launch-your-agent (Apache-2.0). CMA primitive semantics are drawn from the public Claude Managed Agents docs. No upstream code is copied verbatim. License: MIT (this plugin).