Does adding agent instrumentation require changing my agent's code?

Not always. OpenTelemetry’s zero-code instrumentation intercepts calls into popular libraries and frameworks at runtime, through mechanisms like bytecode manipulation or monkey-patching depending on the language, and emits spans for them without touching your application code. For an agent built on a framework or model SDK a zero-code package already covers, that gets you request and response spans for those calls for free. What it doesn’t reach is logic specific to your own agent: a custom retry loop, a hand-rolled planning step, anything that isn’t a call into an instrumented library. Those need a manual span added with the OpenTelemetry SDK. Most production setups end up running both: zero-code for baseline coverage across framework calls, manual spans layered on top for the parts of the agent that are actually yours.