AgentLedger — The Credit Bureau for AI Agents

Trust infrastructure for autonomous AI agents in financial services.

Live demo: elijah3756.github.io/agentledger

MGT 8803 — NLP and GenAI in Finance — Final Project Georgia Institute of Technology — Scheller College of Business — Spring 2026

---

What this is

AgentLedger is the identity, reputation, and trust-scoring bureau for autonomous AI agents transacting in financial services. As AI agents begin executing trades, payments, and other financial transactions on behalf of consumers and enterprises, a critical gap has emerged: the entire financial system runs on identity and trust infrastructure designed for humans (SSN, FICO, KYC/AML) — but no equivalent exists for agents.

AgentLedger fills that gap with four integrated capabilities:

Capability	What it does
Agent Identity	Cryptographic ECDSA credential per agent, anchored to a W3C DID (did:agentledger:…), revocable network-wide.
Trust Scoring	A 300–900 "Agent FICO Score" built from four sub-scores (reliability, accuracy, compliance, security).
Real-time Attestation	Sub-100ms API call that a bank or brokerage makes before executing an agent-initiated transaction.
Anomaly & Revocation	Continuous behavioral monitoring; auto-revocation when drift consistent with compromise is detected.

This repository

This is the prototype deliverable for the MGT 8803 final project. It is a static single-page web application — no build step, no backend, no package manager — that runs the entire AgentLedger simulation in your browser.

agentledger/
├── index.html
├── css/style.css
├── js/
│   ├── app.js            — hash-based router
│   ├── identity.js       — WebCrypto ECDSA credential issuance
│   ├── agent.js          — simulated financial AI agent
│   ├── scoring.js        — Agent FICO score engine
│   ├── anomaly.js        — rolling-window z-score detection
│   ├── attestation.js    — the attestation API
│   ├── registry.js       — in-memory store with localStorage persistence
│   ├── incident.js       — one-shot GenAI incident writer (Claude, fallback)
│   ├── autoresponder.js  — agentic investigator loop (Claude + 6 tools)
│   └── ui/
│       ├── landing.js
│       ├── demo.js       — 7-step guided walkthrough
│       ├── sandbox.js    — interactive dashboard
│       ├── about.js      — architecture + rubric map
│       └── components.js
├── assets/seal.svg
├── README.md
└── LICENSE

Running locally

# From the repo root:
python3 -m http.server 8000
# then open http://localhost:8000

Any static file server works. Because the prototype uses the WebCrypto API, it must be served over http://localhost (or HTTPS in production) — opening index.html directly via file:// may not work in all browsers.

Deploying to GitHub Pages

1. Push this repo to https://github.com/elijah3756/agentledger.
2. In the repo's Settings → Pages, set Source to Deploy from a branch and Branch to main / / (root).
3. Wait ~1 minute; the site will be live at https://elijah3756.github.io/agentledger/.

The .nojekyll file in the repo root disables Jekyll processing so underscore-prefixed files (if any) are served correctly.

Using the GenAI Incident Report Writer

The prototype includes an optional GenAI layer that calls Claude to write plain-English compliance incident reports when an agent is flagged or revoked. Because GitHub Pages has no backend, this is bring-your-own-key:

1. Open the Sandbox, click ⚙ Claude API Key, and paste an Anthropic API key (get one at console.anthropic.com).
2. The key is stored only in this browser's sessionStorage and is cleared when you close the tab.
3. When you click Generate Incident Report in the demo or sandbox, the prototype calls the Anthropic Messages API directly (model: claude-haiku-4-5-20251001).
4. If no key is configured, a deterministic template fallback is used — so the demo never fails.

Rubric coverage

The MGT 8803 final project rubric spans three deliverables (report, pitch deck, prototype) and roughly thirty evaluation dimensions. This prototype addresses the prototype and GenAI dimensions; the companion report and pitch deck address the rest. Traceability:

Rubric dimension	Artifact	Location
Industry dynamics	Report	§2 — Agentic Finance Landscape
Problem / pain points	Report + Prototype	§3; Landing page
Solution & prototype	Prototype	Guided Demo (7 steps) + Sandbox
Target customer	Report	§5.1 — three-tier taxonomy
Market sizing (TAM/SAM/SOM)	Report	§5.2
Customer acquisition	Report	§5.3
Business model & revenue	Report	§6
Competition matrix	Report	§7
Incumbent response	Report	§7
Financial projections	Report	§8
Regulatory environment	Report	§9
SWOT analysis	Report	§10
Porter's Five Forces	Report	§11
Stakeholder impact	Report	§12
18-month roadmap / exit	Report	§13
Technical feasibility	Prototype	Live attestation API, anomaly detection, revocation
GenAI / NLP integration	Prototype	Claude-powered incident report writer (js/incident.js)
Agentic AI	Prototype	Multi-turn Claude agent with tool use — investigates flagged agents and executes real state changes on the registry (js/autoresponder.js)
Pitch deck	Deck	AgentLedger_Pitch_Deck.html

Honest scope — prototype vs. production

The report describes a production architecture (Kafka, Flink, LSTM, Isolation Forest, gradient-boosted trees, Kubernetes, sub-100ms p99 gRPC at 100K QPS). The prototype approximates those components with simpler equivalents that preserve the semantic behavior.

Production	Prototype	Fidelity
PKI with Ed25519 + X.509 chain	WebCrypto ECDSA P-256, browser-issued	Real cryptography
Kafka + Flink aggregation	In-memory registry + localStorage	Semantic equivalent
LSTM + Isolation Forest ensemble	Rolling-window z-scores, 4 dimensions	Captures the pattern
GBDT trust scoring	EWMA over sub-scores	Same monotonic behavior
Sub-100ms p99 gRPC gateway	Direct JS call, latency simulated	Realism without infra
Merkle-tree audit ledger	Append-only localStorage log	Structure, not cryptographic integrity
Claude Opus (server-side)	Claude Haiku (browser, BYOK)	Same family, lower cost

See /about within the running app for the full explanation.

Course and authors

• Course: MGT 8803 · NLP and GenAI in Finance · AI and the Future of Finance
• Term: Spring 2026
• Institution: Georgia Institute of Technology · Scheller College of Business

License

MIT — see LICENSE.