TracePlay: Collaborative Trace Translation for AI Coding Agents
TL;DR
This work presents TracePlay, a research prototype for translating one shared agent trace into role-specific review views while preserving step-level provenance, and suggests that shared step identity, role-specific abstraction, and step-linked discussion are promising design directions for making agent activity more legible and negotiable across stakeholder roles.
Abstract
As LLM-based coding agents become part of software teams, the traces they produce remain easiest for engineers to interpret, leaving product managers and designers without practical ways to inspect implementation decisions. We present TracePlay, a research prototype for translating one shared agent trace into role-specific review views while preserving step-level provenance. Grounded in boundary-object theory [15], TracePlay captures structured trace data via SDK hooks and renders Engineer, PM, and Designer views in a VS Code extension, with a terminal replayer for developer-oriented inspection. We assembled a four-scenario prototype trace corpus (130 steps) and conducted three internal pilot walkthroughs over one OAuth trace to probe cross-role legibility. We treat these walkthroughs as formative evidence: they suggest that shared step identity, role-specific abstraction, and step-linked discussion are promising design directions for making agent activity more legible and negotiable across stakeholder roles.