sunnydubey1111/agent-trajectory-sentinel-demo

AgentTrajectorySentinel Live

A one-class behavioural monitor scoring real agent runs, one step at a time. Pick a run, step through it, watch the score move.

What is real here, and what is not

Real: the runs are committed recordings of a live qwen2.5:7b agent on real tools, with a known failure-onset step. The scores are this project's echo-state-network monitor, fitted on 90 healthy runs from the same corpus. The alarm line (6.4459) is the threshold that spends a 5% false-alarm budget on 30 held-out healthy runs — not a number picked to make the demo look good. The monitor never sees the failure label.

Not real: the scoring happened when this page was built, not when you clicked. Hugging Face charges for Spaces that run Python, so this one is static. Every number came out of the code in the repository below and can be recomputed with py -m devtools.hf_space --build.

What to look for

• —Of the 16 failing runs shipped here, 14 are caught and 2 are missed. The misses are included on purpose: a demo showing only successes is an advertisement.
• —On a healthy run the score should stay flat and under the line.
• —Short runs are hard. This monitor needs a few steps after the onset to accumulate evidence, which is measured in the paper and visible here.

Paper: <https://arxiv.org/abs/2608.02464>

Code and data: <https://github.com/sunnydubey1111/agent-trajectory-sentinel>