Skip to content

Latest commit

 

History

History
71 lines (52 loc) · 3.57 KB

File metadata and controls

71 lines (52 loc) · 3.57 KB

AgentOps — Highlights

A concise tour of what AgentOps is, the decisions behind it, and what it demonstrates. Part of Fischer Product Lab — secure AI systems for trust, risk, and enterprise execution.

Live demo: https://agentops-fpl.vercel.app/


Overview

AgentOps is an enterprise AI agent governance control tower. It inventories the AI agents running across an organization, scores each one's launch readiness with a deterministic rubric, and makes risk impossible to ignore — so leadership can decide what to scale, pause, or govern with evidence rather than instinct.

The problem it solves

Enterprises are deploying AI agents faster than they can govern them — ticket triage, release notes, support summaries, policy Q&A — each spun up by a different team, touching different data, holding different tool permissions. Leadership rarely has a single place to answer the questions that matter: which agents exist, who owns them, what can they touch, and which are actually safe to scale? AgentOps turns that sprawl into one governed, measurable, executive-ready view.

What it does

  • Executive dashboard — portfolio value, readiness mix, hours saved, and a prioritized "needs attention" list.
  • Risk-aware registry — every agent with owner, team, data-access level, autonomy, and readiness, with high-risk attributes flagged before you even open an agent.
  • Agent detail — purpose, data and tools, evaluation metrics, a readiness scorecard with a 7-criteria breakdown and stated reasons, documented failure modes, and an audit log.

The readiness engine (the centerpiece)

Readiness is a pure, deterministic function — explainable and auditable, not a black box:

  1. Seven weighted criteria (0–4 each) normalized to a 0–100 score.
  2. Mapped to a band: Launch / Conditional / Needs Review / Do Not Launch.
  3. Hard safety gates that a high average can never override — if data-safety or tool-permission risk hits the critical floor, the verdict is capped or blocked outright.

The product thesis in one line: a strong agent that is unsafe is still unsafe, and the scoring encodes that judgment so it can't be averaged away. Verdicts are always computed, never stored, so every status is a reproducible function of its inputs.

Engineering & security highlights

  • Secure by design: synthetic data only, read-only (no write paths or admin surface), zero client-side secrets, and no environment variables required in V1.
  • Documented governance: a PRD, a STRIDE threat model, and a security policy.
  • Hardened supply chain: Dependabot, secret scanning, CodeQL, pinned dependency overrides, and a branch-protected pull-request workflow on main.
  • Modern stack: Next.js 16 (App Router) + React 19, TypeScript in strict mode, Tailwind v4, Recharts, Zod, deployed on Vercel with static generation.

What it demonstrates

Product strategy and AI-platform thinking, security and risk judgment, deterministic and explainable decision systems, and executive-grade communication — the ability to turn a messy enterprise workflow into a governed, measurable product.

Roadmap

  • V1.1 — controlled interactivity: filter/search the registry and a live "what-if" rubric slider that recomputes readiness on the fly.
  • V2 — intelligence layer: server-side, deterministic-policy-first AI assist for risk narratives and launch-readiness memos, fully audited and human-in-the-loop.
  • V3 — persistence: synthetic data backed by Postgres with row-level security.