This is my entry for the Sanity Challenge, Path One: Ship an Agent That Queries Real Content. Code: Nuera Quicksilver (MIT). Sanity project ID: f87t11g1. Judge guide · public synthetic demo dataset.
An agent that can propose work, but cannot authorize itself
Nuera Quicksilver is a company-operations console built around a simple boundary: the agent proposes; a deterministic kernel decides what is allowed; a human approves consequential work. The company’s policies, evidence, people, agents, capabilities, objectives, workflows, decisions and metrics live as structured Sanity documents. The agent reads that model through Sanity Context instead of treating a folder of prose as the whole source of truth.
The company model includes typed documents for organizations, departments, entities, capabilities, policies, objectives, workflows, evidence, decisions and metrics. References and fields matter: the kernel can check who holds a capability, which policies apply or supersede one another, whether policies conflict, and what risk follows from an action. Those checks are deterministic TypeScript—not another model prompt.
The assistant uses two Sanity Context read paths: GROQ queries against the live structured dataset, and a Sanity Knowledge Base built from policy and evidence documents. The Knowledge Base contains 12 evidence and policy documents; the judge guide describes how to inspect source-backed answers and the public dataset. The seeded scenario includes two policies in conflict and evidence that challenges the proposed explanation for a production issue. A keyword search can find a sentence; it cannot, by itself, resolve references, scope, supersession, authority and conflicting evidence into a governed decision.
The decision lifecycle is also modeled as content in Sanity. Its states, transitions and structured guards are evaluated by the kernel. The model may recommend an action, but it does not get an approval, execution or publishing tool. Approvals and rollbacks require a human. “Execute” in this demo records a simulated state change and metric; it does not operate production equipment or call external business systems.
Try it
- Live app: project-quicksilver.vercel.app
- Judge guide: FOR-JUDGES.md
- Source: github.com/nueraresearch/project-quicksilver
-
Sanity project:
f87t11g1 -
Public synthetic dataset: query the public
demodataset. The live app uses a privateproductiondataset;democontains synthetic data with the same schema and policy/evidence examples, not customer data.
The live console requires sign-in. For challenge judging, open /sign-in, choose Use an access token instead, and use this judge access token:
qs_mh3a1ExwuX9QYAbsgjAFve6QdXgReucf6bU3Quug8Pk
The token acts as a supervisor in the live deployment. Actions are recorded under its entity in Sanity, so please use it only for challenge evaluation. Execution is simulated; no external system is changed. The judge guide walks through the decisions inbox, the approval boundary, audit trail, chat citations and dataset queries.
What is and is not proven
The repository's credential-free verification command is npm run verify; the README reports 1,282 tests. The live Context endpoints were verified by hand, while the CI tests use fakes for model and Context services. The dedicated project has not had the previous challenge build's live stress test and end-to-end test rerun against it. There have not been task-based usability sessions with real users. No real email or webhook has been sent, and execution is a simulation. Those limits are documented in the judge guide and README.
Top comments (0)