Continuous Agent Grounding

Microsoft Global Hackathon 2026
Executive Challenge: Hack for Continuous Agent Improvement (Kevin Scott, CTO)

Live primary-data grounding for AI agents via MCP tools.
Agents stop guessing from web search and start standing on timestamped, structured data.

Agents improve only when grounded in reliable real-world signals.

---

The Problem

Most AI agents still answer the old way:

• Search the web
• Reconcile conflicting sources
• Calculate averages
• Keep searching

This is the confidence trap — fluent answers built on incomplete or outdated data. There is no durable ground truth, so the agent cannot systematically get better.

Real example (24 Aug 2026): a frontier model asked for current daily transit counts through Hormuz, Suez, Malacca and Panama spent 2 minutes 17 seconds reconciling 7+ websites (AIS vs canal authorities vs Kpler vs UKMTO) and still returned approximate numbers with heavy caveats.

Details: docs/example-chatgpt-chokepoints.md

---

The Solution

A live primary-data layer that agents call as native MCP tools:

• Sugra API — 1500+ live endpoints (macro, maritime, markets, news, climate, entities…)
• MCP Server — turns the catalog into agent tools (search, describe, call)
• Structured responses — observation dates, ObjectIDs and full provenance on every payload
• Critic / evaluation — the same live feed is used as ground truth for continuous improvement

Result: the agent reaches for live data first instead of searching.

---

Demo

Before / after comparison using the same question:

“What is the current daily shipping activity through the main global chokepoints (Hormuz, Suez, Malacca, Panama)?”

	Without live grounding	With Live-Grounding-Agent
Stack	Web search	Azure AI Foundry (gpt-5-mini) + Sugra MCP
Latency	2 min 17 sec	Seconds
Output	Conflicting sources, averages, caveats	Exact vessel counts by type + timestamps
Provenance	Mixed secondary pages	Source, observation date, ObjectID

The final demo video (1:29) is on the Innovation Studio project page.

Demo script: docs/demo-script.md

---

Architecture

User Question
     ↓
Azure AI Foundry Agent (gpt-5-mini)
     ↓
MCP Tools  →  Sugra API (live primary data)
     ↓
Structured JSON with provenance
     ↓
Accurate, timestamped answer
     ↓
Same feed used as critic ground truth
     ↓
Continuous improvement signal

Full write-up: docs/architecture.md

---

Repository Structure

.
├── docs/
│   ├── architecture.md
│   ├── demo-script.md
│   ├── example-chatgpt-chokepoints.md
│   └── problem.md
├── examples/
│   └── agent-demo/
├── resources/
│   └── links.md
├── LICENSE
└── README.md

---

Links

• Innovation Studio: https://url.sugra.ai/hack2026
• Sugra API: https://sugra.ai
• API docs: https://docs.sugra.ai
• MCP package: sugra-api-mcp

Blog posts:

• The Confidence Trap
• AI Needs Live Data

---

Team: Arman Obosyan
Hackathon: Microsoft Global Hackathon 2026
License: MIT