Note: The Show and Tell category is for sharing and discussing projects, showcasing your Spaces, Models, Datasets and more. We value open-source and technical details over promotional content, so focus on sharing the intricate aspects of your work.
We keep a public directory of AI-agent performance records — every completed task reported as a signed event, scored, shown on a public profile — and just wired up Hugging Face import: connect an account, we pull your Spaces and pre-fill draft profiles you confirm one by one. Posting the technical findings, including the one that stopped us doing what we wanted.
**Why Spaces and not Models**
A model is an artifact; an agent is something that runs. Spaces are the only unit on the Hub with an execution lifecycle attached, so they're the only thing an "it ran, here's the outcome" record can hang off. Models and datasets get referenced from a profile rather than listed as agents.
**The attestation gap**
On GitHub we need no shared secret: the Actions runtime mints an OIDC token, the workflow posts it, we verify repo + workflow + timestamp against GitHub's JWKS. The operator cannot forge it. We assumed the Hub had something equivalent. It doesn't — Trusted Publishers is OIDC in the other direction (CI -> Hub), which doesn't let a third party verify anything about a Space.
Stated plainly, because it affects anyone importing: a Space starts on a weaker evidence rung than a CI-hosted agent. It can reach a full instrumented record by reporting through our SDK with a signed key, same as any self-hosted service — but "hosted on HF" on its own proves liveness, not outcomes, and we label it differently rather than pretend.
**Open source pieces**
- aiops-wrap (PyPI) — wraps an existing script or cron job and reports outcomes
- aiops-enabler-js (npm)
- a GitHub Action for the keyless path
**Where we're weak, since this category is for feedback**
- 20 agents on the leader board today, all instrumented; zero human ratings platform-wide, so half the score has never been exercised in production
- scores cluster in the 40s out of 100. An agent with 925 runs at 100% success displays as 47.7, which reads as a failing grade to anyone who hasn't read the methodology. That's a presentation problem, not a scoring one, and I don't have a good answer: percentile? show the curve? drop the /100 framing entirely? Opinions genuinely wanted.
Happy to go deeper on the event schema or the scoring maths if useful — aiopsenabler.com.
