
LangSmith Agent Platform: Why Every Agent Now Gets a Record
Agentic AIWhen a Glean agent’s costs spike, who do you call? LangChain says AI platform teams often can’t answer that, and it has rebuilt LangSmith so every agent, including ones built on other vendors’ platforms, gets a record with a named owner, an ID and a type. For platform leads facing agent sprawl, the change turns observability tooling into an agent inventory.
What Happened
In a post explaining why it rebuilt LangSmith around the agent, LangChain’s Sam Crowder describes the new LangSmith Agent Platform, announced alongside the company’s Series C as “a single place to build, run, improve, and govern every agent.”
The core change is structural. Traces used to sit in a flat project hierarchy with no link to any particular agent, so the same agent’s traces in different environments had no association with each other. Now there is one agent record, with local, dev, staging and production environments grouped under the same agent ID. Its traces, evals, Agent Server deployment and Engine issues are one click away.
Each record also carries a type, which decides its dashboard and evaluation criteria. General purpose and coding agents are available today, with voice “on the horizon.” A coding agent’s dashboard shows merged PRs, token cost and the kind of work it completes, and the coding type can connect a GitHub repo to track downstream outcomes.
Why It Matters
LangChain frames agent sprawl as a governance problem arriving from three directions: engineering teams build custom agents, business units adopt vendor agents such as ServiceNow for IT and Glean for search, and employees pick up coding and productivity agents on their own. Standardizing everyone on one platform “isn’t realistic,” the post argues.
So LangSmith registers agents from any platform that supports the OpenTelemetry Protocol (OTLP), and measures them the same way as agents built in-house. Cost and quality metrics roll up to each agent across every environment. The opinionated dashboards are the other half: because every agent of a given type is measured identically, three engineers running coding agents get comparable numbers without building custom views first.
The approach mirrors AWS’s move toward a central catalog, which Stack Archive covered in the AgentCore agent registry. The difference is that LangSmith anchors the registry in trace data rather than deployment metadata.
What This Means for Builders
LangSmith is composable: teams can run agents elsewhere and connect them only to improve and govern them. Those that build and run on LangSmith get tighter loops, such as LangSmith Engine validating a proposed fix against an Agent Server preview before merge.
The roadmap is where governance gets concrete. Every agent already has a unique, immutable ID; next, LangChain says, agents will be registrable in an identity provider so security teams can manage and audit them like any other identity. Org-level dashboards that aggregate performance across a company are also in progress, beyond today’s workspace-level views. The platform is available in cloud, BYOC and self-hosted environments.
The limits are worth noting. Only two agent types exist, so voice and other specialized agents fall back to the general dashboard for now, and the cross-vendor promise depends on each vendor platform emitting OTLP traces.
The So What
AI platform teams that can’t list their agents and owners should test LangSmith’s registry against their vendor agents first, not just in-house ones: connect one ServiceNow, Glean or coding-agent workload over OTLP and check whether cost and quality roll up to a named owner. If they do, the agent record becomes a practical inventory well before the identity-provider integration ships.
Source: LangChain Blog
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →