Why AI agent observability needs a system boundary

Agent behavior crosses model APIs, orchestration code, shells, package managers, repositories, containers, and remote services. A single application callback rarely owns that whole path.

Separate intent from observed effect.

Identify gaps left by one instrumentation boundary.

Choose complementary telemetry instead of one universal tool.

The semantic gap is the investigation target

A tool-call record says what the agent requested. Process and file evidence says what executed. Connecting the two lets a reviewer distinguish a declared action from the low-level effects it produced.

System evidence does not replace application context

Framework spans, gateway logs, and evaluation platforms remain useful. AgentSight adds another boundary for closed components and host effects, and can export supported model activity into standard telemetry paths.

Primary sources

AgentSight README
OpenTelemetry instrumentation