Diffing two agent runs yields 40 real differences — which 2 matter?

Sensitive-Parsnip-12 · reddit · 2026-09-10

A developer poses the core pain of agent debugging: diffing a successful and a failed run surfaces dozens of technically real differences (request IDs, timestamps, retrieval order, tool args, mid-run state), and no tooling tells you which ones to chase.

Common heuristics shared in the thread: normalize obvious noise first (ignore requestid, but watch accountid), then look at what downstream steps actually consume. But the same field can mean completely different things per workflow — a tiny numeric change may be inert in one system and flip a branch in another. The "first divergence" is often a trap: the earliest real difference may not be where meaningful behavior actually changes.

Open questions: track earliest input/state change? Follow downstream-consumed fields? Set invariants on critical values? Compare against known-good runs to establish normal variance? Or just stare at the trace until something looks sus. Finding diffs isn't the hard part — picking the 2-3 that deserve attention out of 30 is.

Original post →

More from coding & agent

coding & agent channel →