sugeerth.github.io / agentdiff / try

Try AgentDiff in your browser

Two agents ran the same task, and one went wrong. Pick a pair of runs, or drop in two traces of your own, and AgentDiff names the step where they parted, with the evidence for it. The engine runs here, in your browser, as Python compiled to WebAssembly. Nothing you choose is uploaded.

Loading the engine (Python in WebAssembly, about 12 MB the first time)…

A pair from the demo

Pair

The demo agents are simulated. The last pair is real: one agent under two names.

Your own two runs

Run A

Run B

An AgentDiff trace, an OpenTelemetry (OTLP/JSON or GenAI spans) export, an OpenAI or Anthropic message list, or a Claude Code session (.jsonl). The format is detected and named.

What AgentDiff found

	Agent	Outcome	Steps	Tokens	Cost	Read as

The full report is the same page agentdiff compare writes: every step of both runs, the decisive step, why, and what to carry forward.