A Chinese tutorial project that grew production-grade scaffolding

What it does

HelloAgents is a Python multi-agent framework built atop OpenAI-style APIs. It bundles ReAct, reflection, and plan-solve agents with tool registries, session persistence, token counting, circuit breakers, and optimistic file locking. The project explicitly maintains two branches: a stable learn_version tied to the Datawhale tutorial, and a moving main branch where new features land.

The interesting bit

The framework treats “tutorial code” as a first-class release artifact. Every historical release (v0.1.1 through v0.2.9) maps to a specific tutorial chapter, so learners can check out the exact state that matches their progress. Meanwhile the main branch has sprouted enterprise-tinged machinery like TraceLogger, CircuitBreaker, and DevLog decision journaling — a rare case where pedagogical scaffolding and production plumbing share a repo.

Key highlights

• Auto-detects LLM provider from base_url patterns (OpenAI-compatible, Anthropic, or Gemini) without manual adapter selection
• Ships four built-in agent patterns: SimpleAgent, ReActAgent, ReflectionAgent, and PlanAndSolveAgent
• Includes sub-agent delegation via TaskTool and ToolFilter, plus a SkillLoader for external knowledge modules
• Context engineering stack: HistoryManager, TokenCounter, ObservationTruncator, and ContextBuilder
• SSE streaming, async lifecycle hooks, and four logging paradigms baked in
• Community Go port (HelloAgents-go) exists for cross-language learners

Caveats

• CC BY-NC-SA 4.0 license explicitly prohibits commercial use without separate authorization
• The main branch warns that its implementation may diverge from tutorial content; learners must switch branches to stay in sync
• “Production-grade” is the author’s framing; no benchmarks, load tests, or adoption metrics are provided in the README

Verdict

Worth a look if you’re working through the Datawhale agent tutorial and want a reference implementation that doesn’t fall apart when you add real tools. Skip it if you need a battle-tested framework with proven scale, or if your use case is commercial and you don’t want to negotiate a license.