LangSmith

What it is

LangSmith is LangChain Inc.’s commercial platform: tracing, evaluation and deployment for LLM applications and agents, plus — since 2026-03-19 — Fleet, a workspace where people who do not write code build, share and run agents. LangChain, the open-source framework company Harrison Chase started in 2022, raised $125M at a $1.25B valuation in October 2025 (Fortune, TechCrunch). Its about page claims more than 1 billion events a day on LangSmith.

That makes it two products under one bill, and an ops buyer should evaluate them separately. The first answers “what did our agent do and did the last change make it worse?” The second answers “can the RevOps analyst build the vendor-intake agent without filing an engineering ticket, and can IT see what it touched?”

What you are actually buying

Tracing and evals. Every run of an instrumented app lands as a trace tree you can filter, annotate and turn into a dataset, then score with LLM judges, code evaluators or human review queues. Instrumentation is near zero on LangChain or LangGraph, and it also ingests OpenTelemetry, so a non-LangChain stack is not locked out. Traces moved to SmithDB, a Rust store LangChain announced at its Interrupt conference on 2026-05-14, with a P50 trace-tree load of 92 ms.

Engine. Launched in public beta at the same event: it watches production traces, groups failures into issues, drafts evaluators and opens pull requests with proposed fixes. It runs every 6 hours on Plus and costs roughly 7–45 LSU per run (1 LSU = $1). That cost is the reason to schedule it deliberately rather than leave it on by default.

Fleet. Formerly Agent Builder (October 2025). Agents carry their own memory, tools and skills, and come in two identity modes: Claws run on one fixed set of credentials no matter who triggers them, and Assistants act as the person who invokes them, each user authenticating through OAuth. Sharing has three levels — can clone, can run, can edit. Each agent can get its own Slack bot handle, and actions that need sign-off land in an Agent Inbox. Skills follow the Agent Skills spec and can be downloaded with the LangSmith CLI into Claude Code, Cursor or Codex. Since October 2026 every Fleet agent is also exposed as an MCP server and an A2A endpoint, and a separate LangSmith MCP server lets a model read your runs, datasets and experiments.

Pricing

Published at langchain.com/pricing, checked 2026-11-04:

• Developer — $0. One seat, up to 5k base traces a month, then pay as you go; 7 LSU of Fleet usage included. Fleet pauses new runs when the allowance runs out.
• Plus — $39 per seat per month. Up to 10k base traces a month included, unlimited seats at $39 each, 37 LSU of Fleet usage, one free Serverless (Small) deployment, access to Engine.
• Enterprise — custom, annual invoice. Cloud, hybrid (SaaS control plane, your data plane) or fully self-hosted, plus bring-your-own-cloud on AWS (GA) and Azure and Google Cloud (beta) as of 2026-11-03. Custom SSO, ABAC and RBAC.

Base traces keep 14 days; extended traces keep 180 days for an extra fee. The pricing page does not print a per-1k trace rate — overage, Fleet beyond the allowance, deployments and sandboxes all draw down LSUs. Fleet model inference is included in LSUs, so no provider key is needed, and every Fleet run is itself a billed trace. VC-backed startups can apply for up to $10,000 in credits.

The real band: a 12-person team on Plus pays $468/month in seats before a single trace past the included 10k, and Engine alone adds $28–$180 a day at four runs (7–45 LSU each) if left on. Braintrust bills the same 12 people $0 in seats on its $249/month Pro plan, so headcount decides more of this comparison than features do. A two-person team pays $78/month in seats, which is where LangSmith comes out cheaper.

Best for

A platform or RevOps engineering team that already builds on LangChain or LangGraph, wants tracing, evals, deployment and a governed agent workspace from one vendor, and has fewer than about ten people who need seats. The specific Fleet case that works: an ops lead ships a Slack-triggered intake or research agent on Assistant credentials, so the agent can only reach what the asking user can reach, and approvals go to the inbox.

Not for

Teams with 20+ subject-matter reviewers who need accounts to label traces — per-seat billing makes them the dominant cost. Not for a team that has not shipped anything; write 20 test cases in a script first. And not as a Microsoft 365-native agent builder: if your agents live in Teams and SharePoint under Entra governance, Fleet is a second control plane.

Versus the alternatives

• Braintrust — the strongest framework-agnostic rival for evals. Pick it when reviewers outnumber engineers or your stack is not LangChain; it has unlimited users on every plan.
• Langfuse — the fastest-growing entrant and the open-source answer: MIT-licensed, self-hostable, now owned by ClickHouse (January 2026). Pick it when data residency or high volume makes any metered SaaS lose, and you have a team to operate it.
• Arize Phoenix — pick it when you already run Arize for classical ML monitoring.
• For Fleet specifically: Microsoft Copilot Studio when the company runs on Microsoft 365; Dust when non-technical teams want assistants grounded in company knowledge with less configuration; Relevance AI for GTM agent teams. For voice-agent testing, Coval covers what LangSmith does not.

If none fit, a checked-in folder of test cases and a CI job scoring them is the honest fallback for a single feature.

Watch-outs

• The LSU meter covers six products, so the bill is hard to read. Fleet runs, Engine passes, deployments, sandboxes and trace overage all draw from one pool, and every Fleet run also counts as a trace. The 2026-11-02 changelog fixed runs being billed twice. Guard: pull the per-user and per-agent cost endpoints (/v1/fleet/usage/users, /v1/fleet/usage/agents) into a weekly review, set workspace spend limits, and schedule Engine instead of leaving it continuous.
• Claws hold shared credentials. A Claw agent acts with one identity for everyone who can run it, so “can run” access is effectively a grant of that account’s permissions. Guard: default to Assistant mode for anything that touches CRM, mail or contracts, give Claws service accounts scoped to the minimum, and keep “can edit” (which includes inbox review) to named owners.
• The model menu changes under you. Fleet removed Kimi K2.6 in early October 2026 and agents fell back to a default model. Guard: pin the model per agent, keep a small eval dataset per production agent, and re-run it after every changelog entry that touches models.
• Retention is shorter than it looks. Base traces are gone after 14 days, and SaaS long-lived retention was capped at 180 days from 2026-09-14. Guard: upgrade only the traces you need for audit to extended retention, and export anything a legal or compliance review needs past 180 days.