LangChain published a comparison of its three open-source layers on August 6, 2026: LangGraph (runtime), LangChain (framework), and Deep Agents (harness). The post names exact production numbers behind LangChain's internal GTM agent and provides a decision tree for choosing the right abstraction. The key finding: picking the wrong layer costs 20× more per inference run.
LangGraph sits at the bottom: nodes are Python functions, edges are conditional routing logic, state is a TypedDict. Every node transition checkpoints to durable storage, so agents survive restarts and replay from any prior state. Interrupt_before can pause execution at dangerous nodes—database writes, external API calls—and hold for human approval. LangChain sits above LangGraph: a loop abstraction (LLM + tools + middleware hooks) running on LangGraph's engine since October 2025 v1.0. Deep Agents sits at the top: the LangChain loop plus middleware, adding virtual filesystem, subagent spawning, skills (on-demand instruction modules), and cross-run memory. LangChain: "Deep Agents is actually just the core LangChain agent plus middleware."
LangChain's GTM agent runs on Deep Agents in production, handling ~10,000 requests per week across 150 active users. 26% is user-initiated; 74% is ambient (scheduled or event-triggered background work). A majority of production agent traffic now runs autonomous, without human intervention at invocation time. The agent deploys via LangSmith Deployments, handling both interactive and scheduled runs.
Token costs vary sharply by layer. Deep Agents consumed 20× more tokens than LangGraph on a math-plus-research task. Overhead comes from planning (a persistent todo list), virtual filesystem writes, and subagent bootstrapping. At GPT-4o-mini pricing, 10,000 daily executions cost ~$10 on LangGraph and ~$200 on Deep Agents. Deep Agents ran faster in wall-clock time because it parallelizes and batches LLM calls; LangGraph makes sequential calls. Trade-off: high autonomy and lower latency cost more; explicit graphs cost less but add latency.
| Layer | Token use (relative) | Daily cost @ 10k runs | Overhead per query | LLM call pattern | Overhead sources |
|---|---|---|---|---|---|
| LangGraph | 1× (baseline) | ~$10/day | 14 ms | Sequential | — |
| Deep Agents | 20× | ~$200/day | Lower wall-clock (faster) | Parallel / batched | Planning (todo list), virtual filesystem writes, subagent bootstrapping |
Production data backs LangGraph for accuracy-sensitive work. Rexera reduced false positives from 35% (single-prompt LLM) to 2% (LangGraph). AppFolio's Realm-X improved 2× on response accuracy and saves 10+ hours per property manager per week; one feature lifted from 40% to 80% accuracy. Klarna cut resolution time from 11 to 2 minutes, now handles 2.5 million conversations, with projected $40 million profit improvement. LangGraph's orchestration overhead—14ms per query versus LangChain's 10ms—is negligible against these results.
| Customer | Metric | Before | After | Additional impact |
|---|---|---|---|---|
| Rexera | False positive rate | 35% | 2% | — |
| AppFolio Realm-X | Response accuracy (one feature) | 40% | 80% | 2× overall accuracy; saves 10+ hrs/property manager/week |
| Klarna | Resolution time | 11 min | 2 min | 2.5 M conversations handled; projected $40 M profit improvement |
Platform constraints matter. LangGraph Platform doesn't support serverless; Vercel and Cloudflare Workers teams need Mastra or Vercel AI SDK instead. LangGraph Platform costs ~$0.001 per node execution; a five-person team with 500 daily active users pays ~$230/month before LLM costs. AgentExecutor is deprecated and maintenance-only until December 2026; new projects use create_agent() or StateGraph.
Start with Deep Agents for a capable harness with batteries included. Use LangChain for fine-grained tool and context control per step. Use LangGraph directly when your workflow mixes deterministic and agentic steps, requires durable state and human-in-the-loop gates, or needs sub-cent-per-node economics at scale.