Langchain Debater — Development skill for Claude Code

LangChain Debate Agent

A multi-role debate agent built with **LangGraph** + **Claude Sonnet 4.6**. The proposer and critic have **web research tools** (Tavily search + extract) and use them via a hand-rolled ReAct loop — arguments are grounded in live web evidence rather than model memory alone. Dual tracing to **LangSmith** and **Langfuse** captures every tool call, every ReAct iteration, every decision. A personal learning playground for getting hands-on with LangChain / LangGraph, agent tool use, and modern LLM observability.

---

What it does

Given a proposition, it runs a structured Oxford-Union-style debate. Five roles share one Claude model; each has its own system prompt and temperature:

• Setup — initialises the state.
• Moderator (opening) — frames a sharp focus question from the topic.
• Proposer — argues for the proposition, engages directly with the critic's previous turn. Has research tools.
• Critic — argues against, must steelman the proposer's position before attacking. Has research tools.
• Moderator (distillation) — between rounds, distils agreement/disagreement and poses a sharper follow-up.
• Judge — at the end of each round, decides whether to continue or emit a verdict. Honors an optional max_rounds cap.

The graph loops through proposer → critic → moderator-distill → judge until the judge rules or the cap is hit. Inside `proposer` and `critic`, each turn runs a **ReAct tool-call loop**: the LLM can call `tavily_search` / `tavily_extract` up to `MAX_TOOL_CALLS=4` times before producing its final prose. Sources appear inline as markdown links `[claim](https://url)`.

Example output (excerpt)

TOPIC: Is transparent algorithmic curation preferable to editorial human curation...
[MODERATOR, round 0]
Should news platforms be required to replace editorial human curation with
transparent algorithmic curation, given that algorithms can be audited for bias
while human editors cannot be held to th