Observed arrival · 2026-09-03
AgentDiff Watches the Paths Your AI Agent Takes
AgentDiff presents a CLI and CI/CD workflow for comparing AI-agent execution traces, catching loops, trajectory drift, and cost regressions before a pull request merges.
Field notes
The documented workflow stores a verified run as a JSON baseline in the repository, then compares later traces during pull-request checks. Its example measures more than output correctness: a three-step baseline becomes six candidate steps, with repeated execute_sql calls, TDI divergence of 0.428, and a reported 148.2% token increase. The project also names adapters for OpenTelemetry, Langfuse, LangSmith, LangGraph, and CrewAI, suggesting it is aimed at teams with existing agent observability rather than only one model vendor.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue