v0.1.0 — Agents League submission

First public cut of JANUS — the Microsoft Agents League @ AI Skills Fest 2026 submission for the Reasoning Agents track. A decision guardrail that intercepts a proposed agent action and returns a cited, simulated, human-gated recommendation.

What works end to end

• Foundry IQ agentic retrieval — a visible query plan, reranker scores, [ref_id] citations, and abstain-below-floor (no fabricated precedent).
• The headline beat — drag the dependency lever across the 70% concentration knee and the recommendation provably flips MODIFY → APPROVE, because the seeded Monte Carlo + DoWhy do() tail risk genuinely changes. A real causal response, not a scripted animation.
• A real human-in-the-loop gate — the Microsoft Agent Framework workflow pauses at request_info; a server-side run registry resumes the same workflow object on a second request. Idempotent under double-click.
• Measured safety — Content Safety Prompt Shields (direct + indirect/XPIA) and a groundedness gate, with a committed red-team probe (84.6% injection block rate, zero false positives) and a 22-case eval scorecard (4.68/5 groundedness, 4.36/5 relevance).
• Real deploy — azd up provisions Container Apps + a user-assigned managed identity + keyless data-plane roles + Key Vault + App Insights from Bicep. DefaultAzureCredential throughout; no keys in the repo.

Stack

Microsoft Agent Framework · Foundry IQ (Azure AI Search agentic retrieval) · Azure AI Content Safety · Azure OpenAI · FastAPI + Pydantic v2 (uv) · Next.js 15 + Tailwind v4 + React Flow over SSE · OpenTelemetry → Phoenix / Azure Monitor.

Known limits (roadmap, labeled as such)

Work IQ / Fabric IQ as additional sources; durable crash-resumable workflow state; Bayesian simulation; groundedness reasoning-mode (gated by a model deprecation); the cloud AI Red Teaming Agent. One real integration shown working beats three half-wired.