LangGraph Multi-Agent

What this skill does

Build, review, and modernize production-grade multi-agent systems with LangGraph/LangChain while staying version-accurate by default: always resolve the current APIs from docs + installed versions, and fall back to opensrc/ source snapshots for under-the-hood edge cases.

Default operating rules (do these every time)

1. Treat all LangGraph/LangChain APIs as versioned; never “code from memory” for any API surface that might have changed.
2. Establish ground truth before coding:
   • Determine installed/pinned versions (repo lockfiles + importlib.metadata).
   • Query official docs via langchain-docs.SearchDocsByLangChain.
   • Use Context7 for canonical API references and examples.
3. When docs are ambiguous or behavior is subtle, inspect dependency internals via opensrc/ (read-only):
   • Run npx opensrc pypi:<package>@<version> --modify=false
   • Check opensrc/sources.json and cite exact opensrc/... paths + versions in your writeup.

Use references/research_playbook.md when you need a rigorous doc-sweep, including llms.txt-driven crawling of LangGraph docs.

Workflow decision tree

• Build something new → follow Build workflow.
• Audit/deprecations/migration → follow Audit & migrate workflow.
• Latency/cost/scale issues → follow Performance & scale workflow.

Quick start (repo-aware)

1. Generate a deprecations + framework audit report:
   • Run python scripts/audit_repo_agents.py --root . (from the skill folder)
   • Optional: python scripts/audit_repo_agents.py --root . --json agent_audit.json
   • Optional: python scripts/generate_migration_plan.py --audit-json agent_audit.json --out migration_plan.md
2. If you need “latest docs” for a specific area, start with:
   • langchain-docs.SearchDocsByLangChain using queries from references/docs_index.md
3. If you hit a behavior edge case, snapshot internals:
   • Run python scripts/opensrc_snapshot.py --packages langgraph langchain langchain-core (from the skill folder)
4. If you need a doc sitemap for LangGraph (for agentic RAG or doc crawling):
   • Run python scripts/fetch_llms_txt_urls.py --print --unique (from the skill folder)
5. If you need to build an offline docs cache (bounded crawl):
   • Use seeds in references/doc_crawl_targets.md
   • Run python scripts/crawl_docs.py --llms-txt https://langchain-ai.github.io/langgraph/llms.txt --allow-prefixes https://langchain-ai.github.io/langgraph/ --out-dir docs_cache_langgraph (from the skill folder)

Reference map (use these, don’t guess)

• references/langchain_create_agent_middleware.md: create_agent + middleware hooks + migration mappings.
• references/langchain_multiagent_handoffs.md: choosing between subagents vs handoffs vs multi-node subgraphs.
• references/langgraph_graph_api_primitives.md: reducers, Send API, subgraphs, and common error codes.
• references/memory_and_context_engineering.md: state vs store vs runtime context; memory design.
• references/mcp_integration_patterns.md: MCP client/interceptors and ToolRuntime-driven auth/DI.
• references/testing_evaluation.md: tests + eval strategy for safe migrations.
• references/deployment_agent_server.md: langgraph.json and Agent Server deployment basics.
• references/security_threat_model.md: threat model + mitigations for tool calling systems.
• references/upgrades_and_versioning.md: repeatable upgrade process.
• references/audit_and_migration_methodology.md: end-to-end audit→plan→execute methodology.
• references/api_map_python.md: import-path cheat sheet (verify for your version).
• references/ui_nextjs_rsc.md: Next.js App Router (RSC) + React UI integration (Agent Server + useStream).
• references/ui_nextjs_ai_sdk.md: Next.js App Router using AI SDK v6 (useChat) + Streamdown (alternative UI stack).
• references/ui_fastapi_backend.md: FastAPI backend patterns for in-process agents or Agent Server BFF/proxy.
• references/ui_streaming_protocol.md: simple SSE event protocol for custom UI backends.
• references/ui_streamlit.md: Streamlit UI integration patterns (streaming + HITL).

Build workflow (LangChain v1 + LangGraph v1+)

1. Select the right multi-agent topology (start simple):
   • Supervisor + subagents (tool-calling): a main “supervisor” calls subagents as tools for context isolation.
   • Orchestrator-worker (fan-out): use LangGraph’s worker primitives (e.g., Send-style fanout) for parallelizable tasks; use reducers to avoid concurrent state-update errors.
   • Deterministic workflow + agent nodes: keep control-flow explicit; use agents only where needed.
2. Define strict tool schemas and failure behavior:
   • Make tools idempotent where possible; add timeouts and retries; “fail open” on non-critical model/rerank steps.
3. Do context engineering explicitly:
   • Use typed state for dynamic runtime context; keep “store” for long-term memory; use runtime context injection for user/org-scoped dependencies.
4. Add safety controls early:
   • Use middleware guardrails and human-in-the-loop for high-stakes tools (payments, outbound emails, destructive actions).
5. Add observability before scaling:
   • Enable tracing (LangSmith/OpenTelemetry), record tool latency/cost, and add regression evaluations.

Use references/patterns.md for design templates and “gotchas” per topology.

Audit & migrate workflow (deprecated patterns → modern stack)

1. Inventory current architecture (dependencies + imports + runtime behavior):
   • Run scripts/audit_repo_agents.py and expand it with repo-specific patterns if needed.
2. Identify deprecated patterns and their replacements using official docs:
   • Prioritize the LangGraph v1 and LangChain v1 migration guides.
   • Verify whether langgraph.prebuilt.create_react_agent usage should move to langchain.agents.create_agent.
3. Plan the migration in slices (keep it shippable):
   • Preserve external behavior; migrate internals; then upgrade prompts/tools; finally tighten types and add tests.
4. Execute migration and harden:
   • Replace legacy agent frameworks (LlamaIndex agents, CrewAI, Agno, OpenAI Agents) with LangGraph/LangChain equivalents.
   • Add offline tests with fixtures/mocks, and add trace-based regression checks when possible.

Use references/migration_supervisor.md and references/migration_other_frameworks.md for playbooks.

Performance & scale workflow

1. Add measurement first:
   • Capture per-node latency, token usage, tool call counts, cache hit rates, and error taxonomy.
2. Apply the “cheapest win” stack in order:
   • Reduce context size (summaries, retrieval, subagent isolation).
   • Cache tool results (deterministic and keyed); add request coalescing.
   • Parallelize only where state updates are safe (reducers, fanout collection keys).
   • Use smaller/faster models for routing/validation; reserve larger models for synthesis.
3. Constrain worst-case behavior:
   • Recursion limits, max steps, max tool calls, and circuit breakers for flaky dependencies.

Use references/performance_cost.md for checklists and tactics.

Bundled resources in this skill

• scripts/: deterministic helpers (audits, llms.txt extraction, opensrc snapshot runner)
• references/: playbooks + templates for docs research, patterns, migrations, safety, observability, and performance
• assets/: copy/paste templates for Python + UI integrations (Next.js, FastAPI, Streamlit)

Start with references/docs_index.md and references/research_playbook.md.