Boxdawn

Out of the loop.

Open-source tools and research for seeing inside agent systems — and finding the work they do twice.

Multi-agent pipelines, coding agents, tool-calling loops — they repeat work they have already done, and the cost lands on the bill before anyone finds it in the trace. Boxdawn is an open-source tool that finds those repeats deterministically — no LLM judge, nothing leaves your machine — and publishes what the measurements do and do not show. Measured on 16,864 sessions across three corpora — 28 real Claude Code sessions, 6,780 Toolathlon trajectories across 22 frontier models, and 10,056 Exgentic agent-LLM traces.

$ pip install "boxdawn[detect]"

open source · runs locally · deterministic by default · LLM-judge opt-in


96.7%
Context resend
2026-08 · CC session 4130c9a7 · input tokens re-sent (byte-exact)
31.7%
LLM-judge precision
2026-08 · unified n=48 · CI [23.1%, 41.0%] · Amendment v2
3/4
Framework support
2026-08 · Anthropic · LlamaIndex · OpenAI Agents SDK PASS on real workload