Agentic coding tools like Claude Code and Codex are no longer just assistants, they’re becoming part of the software delivery pipeline itself. Yet most teams are still evaluating them on intuition: “this feels faster” or “that seems better,” and companies are incurring huge coding agent bills from unmanaged token usage. Now, creating greener software isn’t only about the systems running in production, but also the creation of those systems.
This talk explores treating token efficiency as a green-software problem you can measure. Coding agents already emit rich OpenTelemetry data, from token usage and cache behaviour to acceptance rates. Our speaker Annie ran an experiment where she tried out different approaches to using Claude Code, and measured which one was most efficient. She’ll share how to instrument agentic workflows with zero overhead, how to read cache-hit rate as a waste signal, and what happens when you compare working styles head to head.