LangSmith runs are application units, not operating-system processes
LangSmith documentation defines a trace as a collection of runs for one application operation, and a run as a span-like unit such as an LLM call, prompt-formatting step, or retrieval call. That structure is intentionally semantic. An operating-system process has a different lifecycle and can outlive, overlap, or contain multiple application-level operations.
Instrumentation gives LangSmith richer application context
Current LangSmith integrations can trace supported LLM providers and agent frameworks automatically, while manual instrumentation can wrap arbitrary functions or code blocks. That makes LangSmith a strong source for the agent graph, intermediate values, metadata, tags, and application-defined steps when you control the tracing boundary.
AgentSight adds execution below or beside the application trace
A shell tool can start tests, compilers, browsers, package managers, or MCP servers whose detailed host behavior is not represented as LangSmith runs unless those components are instrumented. AgentSight can add child-process relationships, file paths, network destinations, and resource activity around the selected task without requiring those descendants to emit LangSmith data.
Evaluation is not a system-profiler feature
LangSmith evaluation spans offline datasets, pre-deployment experiments, and online evaluation of production traces and threads. AgentSight does not provide an equivalent quality-scoring workflow. Use LangSmith or another evaluation system to decide whether outputs meet quality goals; use AgentSight when the diagnosis requires machine-level execution context.
Correlation works best when neither side pretends to be complete
Keep the LangSmith trace ID and application run structure intact. Record the corresponding AgentSight session for selected tasks and correlate by time, command, model/provider request, repository version, or another stable identifier. When the two disagree, inspect the measurement boundary: the application trace may omit a descendant, while the system trace may lack the internal semantic reason for a tool choice.
Choose the smallest layer that answers the question
For framework debugging, prompt/output inspection, feedback, or evaluation, LangSmith is usually the direct tool. For a closed-source CLI, unexplained child process, external path, network destination, or local resource phase, AgentSight is the direct tool. Use both only when the investigation genuinely crosses those layers rather than duplicating every trace by default.