type: decision
status: active
timestamp: 2026-07-02
tags: [stack, ai, orchestration, langchain, langsmith, langgraph, deferred]
status: active
timestamp: 2026-07-02
tags: [stack, ai, orchestration, langchain, langsmith, langgraph, deferred]
LangChain ecosystem — deferred, revisit 2026-10-02
LangChain / LangGraph / LangSmith / integrations. Not adopted, not rejected. Current MCP + skills + AGENTS.md stack covers the same surface. Revisit in 3 months if real gaps surface.
LangChain — deferred
Decision
Not adopting LangChain, LangGraph, LangSmith, or integrations layer. Not rejecting either — set a review checkpoint 2026-10-02.
Why deferred (not rejected outright)
The pitch — chains + tools + memory + tracing — maps 1:1 to what MCP + skills + knowledge/ + Claude Code’s built-in traces already deliver. But:
- Multi-agent orchestration is evolving fast. LangGraph’s graph-based branching MIGHT become useful when subagent orchestration outgrows what
Agenttool +pipeline()primitives cover. - LangSmith’s structured eval framework has no direct MCP equivalent yet. If eval discipline matures, it may fill a real gap.
- The
no-rebuilding-free-softwarerule cuts both ways: adopting LangChain is renting a rebuild of your stack; but skipping observability tools is rebuilding LangSmith yourself with grep-through-transcripts.
Why not adopt now
- Duplicates current stack. Chains ≈ skills. Tools ≈ MCP. Memory ≈
knowledge/. Tracing ≈ CC’s transcript +showThinkingSummaries. Integrations ≈ MCP servers via toolbox. - Framework tax. LangChain-Python conflicts with the TypeScript-first pipeline stack (
pipeline-stack-2026-07-01). LangChain-JS exists but the ecosystem is Python-heavy. - No hit limit. Zero MCP + skills workflows have failed for lack of graph-based branching or LangSmith-style tracing (2026-07-02).
- Platform absorption in progress. Cursor 3.9, Antigravity 2.0, and OpenCode are shipping equivalents natively. Waiting = less lock-in.
Review checkpoint
Revisit 2026-10-02. Trigger a re-evaluation if:
- ≥3 concrete workflows fail because MCP + skills can’t handle branch/loop/merge cleanly
- LangSmith adds an integration OR public consumer that makes cost/benefit clear
- Anthropic / OpenAI / Google ship first-party equivalents that are portable
If none of those, extend deferral another 3 months.
Cross-refs
own-memory-rent-intelligence— the reason MCP + skills is our orchestration layermcp-config-single-source-of-truth— the mechanismno-rebuilding-free-software— the constraint- Source: user Q 2026-07-02 after Cracking the Cryptic-style LangChain marketing video