LangChain updated LangSmith: on-call copilot for alert triage, voice debugging, and Deep Agents Rubrics

LangChain published its June digest, which includes key updates to the LangSmith platform, a new system for evaluating agents called Deep Agents Rubrics, an educational course on deployment, and announcements of in-person events for the second half of the year.

Fleet On-Call Copilot for Alerts

The central LangSmith update in June is the Fleet on-call copilot, a tool for rapid triage of alerts in agent systems. The tool's purpose is to help the on-duty engineer quickly get to the root of an incident: it collects context from traces, analyzes error patterns, and suggests the most likely cause of failure without the need to manually review hundreds of lines of logs. This problem is well-known to any team running agents in production: the more complex an agent system is, the harder it is to understand exactly where something went wrong.

An error could have occurred in the prompt, in a tool, in the long-term memory mechanism, or in the orchestration logic of multiple agents at the same time. The Fleet on-call copilot narrows down the search space and reduces the time from the first alert to understanding its cause.

New Features for Developers

In addition to the on-call copilot, LangSmith received four practical additions:

• Computer use for agents — agents can now interact with the computer interface directly: click buttons, fill in form fields, read screen contents in real time
• Voice debugging of traces — developers can play back voice interactions directly in the LangSmith interface and clearly see where the agent made an error in speech interpretation
• Experiment status tracking — a new dashboard shows the progress of long-running test runs without the need for manual checking of each execution
• Programmatic sub-agents — the ability to run nested agent chains directly from code without manual configuration of the orchestration layer

All four updates address specific engineering needs and emerged from requests by teams already operating agent systems in real production, not just prototyping.

Deep Agents Rubrics and Training

LangChain announced Deep Agents Rubrics — a system of structured criteria for evaluating complex multi-step agents. This addresses a long-standing industry problem: how to objectively measure the quality of an agent that solves a non-linear task in multiple steps? Simple metrics like 'success/failure' are not sufficient here.

Rubrics offer a way to break down complex scenarios into assessable sub-tasks and assign scores for each one. This is especially useful when comparing different versions of an agent or when choosing between architectural approaches — for example, between a single large orchestrator agent and several specialized executor agents. Structured evaluation helps make engineering decisions based on evidence rather than relying on the subjective feeling of 'it seems to work better'.

A new course on deploying LangSmith has been released for those who want to move from the first prototype to full production deployment. In-person events in Chicago, Berlin, Washington, and Las Vegas are scheduled for the second half of 2026.

What This Means

The June LangSmith updates reflect an important shift: the industry's center of gravity is moving from agent development to their reliable operation. The on-call copilot, voice tracing, and evaluation rubrics are no longer prototyping tools, but operational infrastructure for teams running agents in real production. AI agents are transitioning from the experimental stage to the stage of engineering discipline.