LangSmith

Fleet agents act through built-in integrations such as Gmail, Slack and GitHub, custom MCP servers added to the workspace, an Arcade integration for third-party tools, and a Salesforce integration that queries records; each agent authenticates either with its own fixed credentials or with each user's own OAuth connection, and admins can enable or disable integrations and control who may invoke each one.

SourceLangChain, docs.langchain.com/langsmith fleet/access-and-oversight and docs.langchain.com/llms.txt (Fleet tools, Arcade, Salesforce and agent identity pages)read 2026-09-21

Deployed agents run as durable Agent Servers in LangSmith Deployment, with background runs, cron, interrupts and double-texting handling. Deployment revisions are kept and agent graphs are debugged in LangSmith Studio, so orchestration is versioned and debuggable. The workflows that mix deterministic nodes with agent steps and make loops, branches and retries first-class are written in LangGraph rather than in LangSmith, and Fleet builds agents from natural-language instructions rather than workflow definitions.

SourceLangChain, docs.langchain.com/langsmith/llms.txt (Agent Server core capabilities, cron and interrupt pages) and langchain.com/pricingread 2026-09-21

Context Hub stores and versions agent instructions, skills and durable memory, Fleet agents read the customer's systems through integrations at run time, and LangSmith evaluates retrieval inside the customer's own application. No document ingestion, index, retrieval layer or knowledge API that grounds agents in company data is documented.

SourceLangChain, docs.langchain.com/langsmith/llms.txt (Context Hub pages) and docs.langchain.com/llms.txt (Fleet section)read 2026-09-21

Fleet agents can be configured to pause and request approval before taking specific actions, and a central inbox lets reviewers approve, edit or reject those actions across all agents; the Agent Server also documents human-in-the-loop and time travel through its API, and admins gate which tools each role may invoke.

SourceLangChain, docs.langchain.com/langsmith fleet/access-and-oversight and llms.txt (human-in-the-loop using server API)read 2026-09-21

Beyond holding SOC 2 Type II, HIPAA and GDPR, LangSmith supports SAML or OIDC single sign-on with just-in-time provisioning and SCIM for automated provisioning and deprovisioning, applies built-in or custom roles per workspace and attribute-based access policies that can block PII from specific users, and records administrative actions in audit logs; retention periods and deletion by metadata are configurable, and LangSmith states it does not use customer data to train models. Fleet adds role- and attribute-based control over which users may invoke each MCP server and integration.

SourceLangChain, docs.langchain.com/langsmith enterprise and fleet/access-and-oversight, and langchain.com/pricingread 2026-09-21

LangSmith traces capture each agent run step by step, including tool calls, decisions and outputs, and can be inspected, searched and exported, with bulk data export on every plan; for Fleet agents, the trace shows which agent acted, on whose behalf and with what credentials. Audit logs record administrative actions across the organization separately from traces, and traces are kept on base or extended retention tiers.

SourceLangChain, docs.langchain.com/langsmith fleet/access-and-oversight, enterprise and llms.txt (audit logs), and langchain.com/pricingread 2026-09-21

A managed deep agent's conversational memory is scoped to its thread, and durable memory, when enabled, gives the deployment a read and write tree in Context Hub that the agent maintains itself, with a hot file loaded into every run and cold files read when relevant; the documentation says what should never be stored there, such as personal data and credentials, and Context Hub contents can be pushed, pulled, listed and deleted through the SDK. LangSmith Deployment's API also carries state and memory endpoints.

SourceLangChain, docs.langchain.com/langsmith managed-deep-agents-memory and llms.txt (Context Hub SDK), and langchain.com/pricingread 2026-09-21

LangSmith runs in LangChain's managed cloud with US or EU data residency; hybrid keeps the control plane in LangChain's cloud and the data plane in the customer's VPC; self-hosting runs the whole platform on the customer's Kubernetes, with guides for AWS, Azure and GCP; and Bring Your Own Cloud stores data and runs services in the customer's cloud while LangChain operates, scales and upgrades the infrastructure.

SourceLangChain, docs.langchain.com/langsmith enterprise and llms.txt (BYOC, hybrid and self-hosted pages), and langchain.com/pricingread 2026-09-21

Fleet gives access to prebuilt agent templates that a team customizes for tools, prompts and channels and can share and clone, and LangSmith Engine is a ready-made agent that monitors traces on its own, clusters agent failures into issues, diagnoses them and recommends fixes.

SourceLangChain, langchain.com/pricing and docs.langchain.com/llms.txt (Fleet templates and quickstart pages)read 2026-09-21

LangSmith Deployment includes cron scheduling for deployed agents, Fleet agents run on recurring schedules and start automatically from configured channels, including Slack and Microsoft Teams apps where the team tags the agent directly, Managed Deep Agents declare cron schedules and start runs from Slack messages, and Fleet agents can also be triggered through the API. Schedules and channel events start the agent without a person initiating it.

SourceLangChain, langchain.com/pricing and docs.langchain.com/llms.txt (Fleet schedules, channels, Slack and Teams pages; Managed Deep Agents schedules and channels)read 2026-09-21

The LangSmith LLM Gateway sits between agents and model providers with the customer's own model API keys or Gateway Credits, model fallbacks, cost and rate-limit controls, and PII and secrets redaction; spend, rate-limit and per-customer policies are documented, and deployments can connect to custom models and OpenAI-compatible providers. Model choice stays with the customer, with policies enforced across it.

SourceLangChain, langchain.com/pricing and docs.langchain.com/langsmith/llms.txt (LLM Gateway fallbacks and policy pages)read 2026-09-21

LangSmith documents a REST API across traces, datasets, evaluators, prompts, deployments, access policies and SCIM, Python and TypeScript SDKs, a LangSmith MCP server and a remote MCP endpoint, and an Agent Server whose deployed agents stream, accept webhooks and can be exposed as MCP servers or reached over A2A; Fleet agents can be invoked through the LangGraph SDK or REST API, and Fleet webhooks connect publishing to CI/CD pipelines. MCP is both served and consumed, and the platform fits into CI/CD and other agent systems.

SourceLangChain, docs.langchain.com/langsmith/llms.txt and docs.langchain.com/llms.txt (Fleet section), and langchain.com/pricingread 2026-09-21

Offline evaluations run against collected datasets and online evaluators run on production traffic, with code, LLM-as-a-judge, composite and multi-turn online evaluators documented, annotation queues for human feedback, a Prompt Hub and Playground for testing changes, and LangSmith Engine, which monitors traces, clusters failures into issues, recommends fixes to prompts and code, and creates offline and online evals so failures do not repeat. That covers testing on datasets before production, quality scored over time, and a loop after deployment.

SourceLangChain, langchain.com/pricing feature table and docs.langchain.com/langsmith/llms.txt (online evaluator pages)read 2026-09-21

Fleet's computer use attaches an isolated virtual computer on which the agent writes and executes code, manages files, installs packages and calls authenticated APIs through a credential-injecting proxy, and LangSmith Sandboxes run agent-generated code in ephemeral environments. Neither documents the agent operating a browser, a desktop session or a graphical interface, so the work stays in sandboxes, terminals and code execution.

SourceLangChain, docs.langchain.com/langsmith/fleet/computer-use and langchain.com/pricing; docs.langchain.com/langsmith/fleet/computer-use.mdread 2026-09-21