Make your best LLM the boss and your cheap ones the interns

What it does

Maestro is a hierarchical task runner written in Python. You feed it an objective, and an orchestrator model—originally Claude 3 Opus, now swappable via LiteLLM—chops the work into subtasks. A cheaper subagent model executes each step with memory of prior results, and the orchestrator refines the batch into a final deliverable. The loop halts only when the orchestrator emits the exact phrase “The task is complete:”. Everything is logged to a Markdown file.

The interesting bit

The project is refreshingly blunt about what it is: a script, not a platform. It uses literal string matching as a termination condition, which feels almost charmingly low-tech next to the agent frameworks shipping this month. The real utility is model-tier arbitrage—using LiteLLM to mix and match orchestrators, workers, and refiners across Anthropic, OpenAI, Gemini, Groq, or a local Ollama instance.

Key highlights

• Supports multiple providers (Anthropic, OpenAI, Gemini, Cohere, Groq) and local runners (Ollama, LMStudio) through LiteLLM integration.
• Iterative loop where the orchestrator assesses its own progress and emits a hardcoded completion phrase to break the cycle.
• Subagents receive memory of previous tasks for context, so later steps build on earlier ones.
• Optional Tavily search integration lets the orchestrator look up context before assigning a subtask.
• Outputs a full exchange log as a Markdown file and can generate code files or folders during execution.

Caveats

• The README is mostly installation quickstart; the actual orchestration logic and prompt engineering details are not shown, so it is unclear how reliably the “task is complete” heuristic works in practice.
• Multiple entrypoint scripts (maestro.py, maestro-anyapi.py, maestro-gpt4o.py, etc.) suggest the codebase is a collection of slightly divergent examples rather than a unified abstraction.

Verdict

Worth a look if you want a minimal, hackable template for tiered LLM delegation without adopting a heavy agent framework. Skip it if you need robust error handling, retry logic, or production observability—the README does not suggest those exist.