diff --git a/src/adapters/langgraph.ts b/src/adapters/langgraph.ts
new file mode 100644
index 0000000..010a7b6
--- /dev/null
+++ b/src/adapters/langgraph.ts
@@ -0,0 +1,187 @@
+import { resolve, join } from 'node:path';
+import { loadAgentManifest, loadFileIfExists } from '../utils/loader.js';
+import { loadAllSkills, getAllowedTools } from '../utils/skill-loader.js';
+
+/**
+ * Export a gitagent directory to a LangGraph-compatible Python agent graph.
+ *
+ * Generates a ready-to-run Python file implementing a stateful ReAct agent
+ * using LangGraph's StateGraph with ToolNode. Supports multi-agent topologies
+ * via the gitagent `agents` manifest field. Compliance constraints are
+ * injected as system prompt rules.
+ */
+export function exportToLangGraph(dir: string): string {
+  const agentDir = resolve(dir);
+  const manifest = loadAgentManifest(agentDir);
+
+  const systemPrompt = buildSystemPrompt(agentDir, manifest);
+  const tools = buildToolDefinitions(agentDir);
+  const hasSubAgents = manifest.agents && Object.keys(manifest.agents).length > 0;
+  const modelImport = resolveModelImport(manifest.model?.preferred);
+
+  const lines: string[] = [];
+
+  lines.push('"""');
+  lines.push(`LangGraph agent graph for ${manifest.name} v${manifest.version}`);
+  lines.push('Generated by gitagent export --format langgraph');
+  lines.push('"""');
+  lines.push('');
+  lines.push('from typing import Annotated, TypedDict, Sequence');
+  lines.push('from langchain_core.messages import BaseMessage, HumanMessage, AIMessage');
+  lines.push('from langchain_core.tools import tool');
+  lines.push(`${modelImport}`);
+  lines.push('from langgraph.graph import StateGraph, END');
+  lines.push('from langgraph.prebuilt import ToolNode');
+  lines.push('import operator');
+  lines.push('');
+
+  lines.push('# --- Agent State ---');
+  lines.push('class AgentState(TypedDict):');
+  lines.push('    messages: Annotated[Sequence[BaseMessage], operator.add]');
+  lines.push('');
+
+  if (tools.length > 0) {
+    lines.push('# --- Tools ---');
+    for (const t of tools) {
+      const funcName = t.name.replace(/[^a-zA-Z0-9]/g, '_');
+      lines.push('');
+      lines.push('@tool');
+      lines.push(`def ${funcName}(input: str) -> str:`);
+      lines.push(`    """${t.description}"""`);
+      lines.push(`    # TODO: implement tool logic`);
+      lines.push(`    raise NotImplementedError("${funcName} not yet implemented")`);
+    }
+    lines.push('');
+    lines.push(`tools = [${tools.map(t => t.name.replace(/[^a-zA-Z0-9]/g, '_')).join(', ')}]`);
+  } else {
+    lines.push('tools = []');
+  }
+
+  lines.push('');
+  lines.push('# --- Model with tool binding ---');
+  lines.push(`llm = ${resolveModelInstantiation(manifest.model?.preferred)}`);
+  lines.push('llm_with_tools = llm.bind_tools(tools)');
+
+  lines.push('');
+  lines.push('SYSTEM_PROMPT = """' + systemPrompt.replace(/"""/g, '\\"\\"\\"') + '"""');
+
+  lines.push('');
+  lines.push('# --- Graph Nodes ---');
+  lines.push('def agent_node(state: AgentState) -> dict:');
+  lines.push('    """Main reasoning node."""');
+  lines.push('    messages = state["messages"]');
+  lines.push('    from langchain_core.messages import SystemMessage');
+  lines.push('    if not any(isinstance(m, SystemMessage) for m in messages):');
+  lines.push('        messages = [SystemMessage(content=SYSTEM_PROMPT)] + list(messages)');
+  lines.push('    response = llm_with_tools.invoke(messages)');
+  lines.push('    return {"messages": [response]}');
+  lines.push('');
+  lines.push('tool_node = ToolNode(tools)');
+  lines.push('');
+  lines.push('def should_continue(state: AgentState) -> str:');
+  lines.push('    last = state["messages"][-1]');
+  lines.push('    if hasattr(last, "tool_calls") and last.tool_calls:');
+  lines.push('        return "tools"');
+  lines.push('    return END');
+
+  lines.push('');
+  lines.push('# --- Build Graph ---');
+  lines.push('workflow = StateGraph(AgentState)');
+  lines.push('workflow.add_node("agent", agent_node)');
+  lines.push('workflow.add_node("tools", tool_node)');
+  lines.push('workflow.set_entry_point("agent")');
+  lines.push('workflow.add_conditional_edges("agent", should_continue)');
+  lines.push('workflow.add_edge("tools", "agent")');
+  lines.push('graph = workflow.compile()');
+
+  if (hasSubAgents) {
+    lines.push('');
+    lines.push('# --- Sub-Agents ---');
+    lines.push('# Sub-agents defined in agent.yaml can be modelled as additional');
+    lines.push('# StateGraph nodes with supervisor routing.');
+    lines.push('# See: https://langchain-ai.github.io/langgraph/tutorials/multi_agent/multi-agent-collaboration/');
+    for (const [name] of Object.entries(manifest.agents ?? {})) {
+      lines.push(`# Sub-agent: ${name}`);
+    }
+  }
+
+  lines.push('');
+  lines.push('if __name__ == "__main__":');
+  lines.push(`    print("Agent: ${manifest.name} v${manifest.version}")`);
+  lines.push('    while True:');
+  lines.push('        user_input = input("You: ").strip()');
+  lines.push('        if not user_input or user_input.lower() in ("exit", "quit"):');
+  lines.push('            break');
+  lines.push('        result = graph.invoke({"messages": [HumanMessage(content=user_input)]})');
+  lines.push('        last_ai = next((m for m in reversed(result["messages"]) if isinstance(m, AIMessage)), None)');
+  lines.push('        if last_ai:');
+  lines.push('            print(f"Agent: {last_ai.content}")');
+
+  return lines.join('\n');
+}
+
+function buildSystemPrompt(agentDir: string, manifest: ReturnType<typeof loadAgentManifest>): string {
+  const parts: string[] = [];
+
+  const soul = loadFileIfExists(join(agentDir, 'SOUL.md'));
+  if (soul) parts.push(soul);
+
+  const rules = loadFileIfExists(join(agentDir, 'RULES.md'));
+  if (rules) parts.push(`## Rules\n${rules}`);
+
+  const skillsDir = join(agentDir, 'skills');
+  const skills = loadAllSkills(skillsDir);
+  for (const skill of skills) {
+    const allowedTools = getAllowedTools(skill.frontmatter);
+    const toolsNote = allowedTools.length > 0 ? `\nAllowed tools: ${allowedTools.join(', ')}` : '';
+    parts.push(`## Skill: ${skill.frontmatter.name}\n${skill.frontmatter.description}${toolsNote}\n\n${skill.instructions}`);
+  }
+
+  if (manifest.compliance) {
+    const c = manifest.compliance;
+    const constraints: string[] = ['## Compliance Constraints'];
+    if (c.communications?.fair_balanced) constraints.push('- All outputs must be fair and balanced (FINRA 2210)');
+    if (c.communications?.no_misleading) constraints.push('- Never make misleading or promissory statements');
+    if (c.data_governance?.pii_handling === 'redact') constraints.push('- Redact all PII from outputs');
+    if (c.supervision?.human_in_the_loop === 'always') constraints.push('- All decisions require human approval');
+    if (manifest.compliance.segregation_of_duties) {
+      const sod = manifest.compliance.segregation_of_duties;
+      if (sod.conflicts) {
+        constraints.push('- Segregation of duties conflicts:');
+        for (const [a, b] of sod.conflicts) {
+          constraints.push(`  - "${a}" and "${b}" may not be held by the same agent`);
+        }
+      }
+    }
+    if (constraints.length > 1) parts.push(constraints.join('\n'));
+  }
+
+  return parts.join('\n\n');
+}
+
+interface ToolDef {
+  name: string;
+  description: string;
+}
+
+function buildToolDefinitions(agentDir: string): ToolDef[] {
+  const skills = loadAllSkills(join(agentDir, 'skills'));
+  return skills.map(s => ({
+    name: s.frontmatter.name,
+    description: s.frontmatter.description ?? s.frontmatter.name,
+  }));
+}
+
+function resolveModelImport(model?: string): string {
+  if (!model) return 'from langchain_openai import ChatOpenAI';
+  if (model.startsWith('claude')) return 'from langchain_anthropic import ChatAnthropic';
+  if (model.startsWith('gemini')) return 'from langchain_google_genai import ChatGoogleGenerativeAI';
+  return 'from langchain_openai import ChatOpenAI';
+}
+
+function resolveModelInstantiation(model?: string): string {
+  if (!model) return 'ChatOpenAI(model="gpt-4o", temperature=0.3)';
+  if (model.startsWith('claude')) return `ChatAnthropic(model="${model}", temperature=0.3)`;
+  if (model.startsWith('gemini')) return `ChatGoogleGenerativeAI(model="${model}", temperature=0.3)`;
+  return `ChatOpenAI(model="${model}", temperature=0.3)`;
+}
\ No newline at end of file
diff --git a/src/runners/langgraph.ts b/src/runners/langgraph.ts
new file mode 100644
index 0000000..080e6e2
--- /dev/null
+++ b/src/runners/langgraph.ts
@@ -0,0 +1,40 @@
+import { writeFileSync, unlinkSync } from 'node:fs';
+import { join } from 'node:path';
+import { tmpdir } from 'node:os';
+import { spawnSync } from 'node:child_process';
+import { randomBytes } from 'node:crypto';
+import { exportToLangGraph } from '../adapters/langgraph.js';
+import { AgentManifest } from '../utils/loader.js';
+import { error, info } from '../utils/format.js';
+
+interface RunOptions {
+  prompt?: string;
+}
+
+export function runWithLangGraph(agentDir: string, _manifest: AgentManifest, _options: RunOptions = {}): void {
+  const script = exportToLangGraph(agentDir);
+  const tmpFile = join(tmpdir(), `gitagent-langgraph-${randomBytes(4).toString('hex')}.py`);
+
+  writeFileSync(tmpFile, script, 'utf-8');
+
+  info(`Running LangGraph agent from "${agentDir}"...`);
+  info('Make sure langgraph, langchain-core, and a model package are installed.');
+
+  try {
+    const result = spawnSync('python3', [tmpFile], {
+      stdio: 'inherit',
+      cwd: agentDir,
+      env: { ...process.env },
+    });
+
+    if (result.error) {
+      error(`Failed to run Python: ${result.error.message}`);
+      info('Install: pip install langgraph langchain-core langchain-openai');
+      process.exit(1);
+    }
+
+    process.exit(result.status ?? 0);
+  } finally {
+    try { unlinkSync(tmpFile); } catch { /* ignore */ }
+  }
+}
\ No newline at end of file