agent-orchestrator

agent-orchestrator

Autonomous meta-agent orchestrator plugin for Claude Code.

Coordinates a team of specialized agents to solve complex tasks end-to-end:

1. Research — 5 Haiku brainstormers explore the problem space in parallel
2. Planning — NotebookLM RAG builds a knowledge base; strategic plan with quality criteria
3. Execution — Codex instances implement subtasks via codex-swarm
4. Quality Gate — each result evaluated against criteria (accept / re-dispatch up to 3x / Opus fallback)
5. Synthesis — results merged, 3-round self-critique produces the deliverable

Install

# One-time: register the marketplace
claude plugin marketplace add dynamic-dome/agent-orchestrator-plugin

# Install the plugin (user scope is typical for personal tooling)
claude plugin install agent-orchestrator@agent-orchestrator-marketplace --scope user

Verify:

claude plugin marketplace list
# → should list "agent-orchestrator-marketplace"

Usage

Invoke the skill by name or via trigger phrases. In Claude Code:

Nutze den agent-orchestrator Skill für folgende Aufgabe: …

Or drop any of the trigger phrases in a message — the skill will activate:

"orchestriere das", "grosses Projekt", "full auto", "lass die Agents los", "ich brauch ein Team dafuer", "recherchiere und baue", "mach das komplett autonom"

Then describe your goal. The orchestrator runs all 5 phases autonomously.

When This Is (And Isn't) The Right Tool

Use it when:

• A task needs real research + planning + implementation (not just one of these)
• You want multiple perspectives (analytic, creative, critical) on a problem
• The task is too large for a single agent pass
• You want a polished deliverable, not a quick answer

Skip it when:

• Quick fix / one-line answer / single file edit — overkill
• You already know the solution and just need it typed
• The research phase would not add value (e.g. "format this JSON")

Dependencies

The orchestrator invokes these at runtime and degrades gracefully if they are missing:

Dep	Purpose	Fallback
multi-model-orchestrator plugin (codex-swarm skill)	Codex execution swarm in Phase 3	Agent tool with model: sonnet
notebooklm user-skill (Python API)	Knowledge base + RAG in Phase 1+2	Inline context kept locally
research-pipeline skill	Web research in Phase 1	WebSearch direct
Agent tool (built-in)	Haiku brainstormers in Phase 1	— (required)

Important: use the notebooklm user-skill (notebooklm-py, Python API), not the deprecated notebooklm Chrome-MCP plugin. The plugin was archived 2026-04-24.

Quality Standards

• Research phase: iterate until blind spots closed (min 3 refinement rounds)
• Planning phase: each subtask must have measurable acceptance criteria
• Quality gate: configurable threshold, default 9.0/10 with 3 re-dispatches, then Opus fallback
• Synthesis: 3-round self-critique before presenting to user

License

MIT — see LICENSE if present.