LangSmith

LangSmith Platform is LangChain Inc.'s commercial product, positioned as a unified agent platform covering the entire AI agent lifecycle: from observability and evaluation through deployment, autonomous improvement (Engine), no-code agents (Fleet), to isolated execution environments (Sandboxes). It is built on the thesis that agent applications need a different kind of infrastructure than classic software: cycles, tools, memory, streaming, and probabilistic outputs require a purpose-built stack. Framework-agnostic — it works with OpenAI, Anthropic, Google, Vercel AI SDK, LlamaIndex, LangGraph, and custom stacks. Supports OpenTelemetry.

Six sub-products. (1) Observability — real-time tracing of every LLM/tool/agent call, live monitoring of latency P50/P99, cost, error rates, token usage; annotation queue with human feedback; unsupervised topic clustering and executive summaries as Insights (beta); alerts via webhooks and PagerDuty. Base traces (14-day retention) and extended traces (180 days). SmithDB — a dedicated database for agent observability, offering significantly faster full-text search and trace queries than generic OLAP.

(2) Evaluation — online and offline evals, dataset collection from production traces, versioned datasets, Prompt Hub and Playground for fast iteration, LLM-as-judge, code-based scorers, human feedback loop. (3) Deployment — 1-click deploy for LangGraph agents (Serverless small/medium/large + Dedicated), Assistants API with 30+ endpoints (state, memory, cron scheduling), streaming intermediary steps, agent authorization (beta), expose as MCP server, LangSmith Studio for visual graph-state inspection.

(4) Engine — autonomous trace monitor running every 6 hours: clusters behaviour into issues, diagnoses failures, generates prompt and code fix suggestions, creates datasets for human review, builds offline and online evals that block regressions. Metered in LCUs (5-30 LCUs per run depending on complexity). (5) Fleet — no-code agent builder for non-programmers (business users), 3 model tiers Fast/Pro/Max (LangChain manages model access, no customer API key needed), remote MCP server tools, prebuilt templates, trigger via API. All runs automatically traced. (6) Sandboxes — ephemeral isolated Linux environments for agent-generated code, per-second billing, auth proxy with custom callbacks, snapshot and fork, bring-your-own-image, port tunneling, sandbox CLI.

Pricing model: seat + usage-based. Developer $0/seat (5k base traces/mo, 1 seat, community support). Plus $39/seat/month (10k base traces, unlimited seats, 1 free Serverless Small, email support, access to Engine/Fleet/Sandboxes). Enterprise custom (self-hosted and hybrid, custom SSO/RBAC, SLA, deployed engineers). Two units: LCU (LangChain Compute Unit, $1.50) and LSU (LangChain Storage Unit, $1.00). LangSmith for Startups offers discounted rates for VC-backed AI companies.

Enterprise customers: LinkedIn, Uber, Klarna, Vanta, Clay, Rippling, Lyft, Gong, Harvey, Abridge, Cloudflare, Bristol Myers Squibb, Workday, Cisco, Mercor, Monday, Nvidia, Bridgewater, Coinbase, Elastic, ServiceNow. Competition: Braintrust (closed-source SaaS), DeepEval / Confident AI (open-source alternative), Weights & Biases Weave, Arize AI, Fiddler AI, Langfuse. LangSmith's differentiator: tight integration with the LangGraph runtime (one-click deploy of the same graph you tested locally) and a unified surface for obs+evals+deployment (one dashboard instead of stitching several tools).