Earshot

The research confirms mature trace foundations for generative systems, agent-specific span conventions, commercial observability platforms and voice-testing demand in another buyer segment. It found no reviewed product applying cited voice narration to coding-agent traces. The buyer, recent timing trigger and exact distinction from adjacent control planes still require validation.

Earshot should separate declared task, source context, observed trace, agent-produced rationale, tool result, diff and test evidence, expected constraint, divergence signal, interpretation-drift candidate, human finding, narration, intervention decision, command, acknowledgment, readback and engineering outcome. A successful tool call does not prove the right objective, and a divergence signal does not prove what an agent understood.

Collection must be allowlisted and minimized; raw prompts, source code, secrets, customer data and private developer activity cannot flow into shared narration by default. The product cannot score workers, assign blame, claim causal diagnosis or use voice as the only accessible surface.

Engineering and developer-productivity leaders at small and midsize software organizations operating coding agents on consequential shared work.