Enterprise Agent Workshop

A hands-on workshop for engineering teams building production-grade Agents inside the enterprise. The samples in this repository walk you from a single-process skill-testing prototype, through multi-agent sandboxed pipelines, all the way to a hosted agent deployed on Microsoft Foundry with full Azure infrastructure.

Designed as a one-day workshop (≈ 6–8 hours of code + slides).

The workshop is organized around four recurring engineering concerns that every enterprise Agent program eventually has to solve:

---

Workshop Modules

1. Skill_Testing_Agent/ — Adversarial Skill Testing 101

A minimal Microsoft Agent Framework + GitHub Copilot SDK (Python) project that tests a “business Agent” (an educational-video-script generator with a strict output template) against an adversarial test Agent. Two models — Claude Opus 4.7 and GPT-5.5 — are run in parallel and compared in the console.

Key concepts introduced:

• A business Agent with a deterministic output contract.
• An adversarial test Agent that crafts edge-case prompts (single-turn and multi-turn).
• A deterministic validator that checks the contract without an LLM.
• An LLM-as-judge rubric scoring the response on 5 dimensions.
• 10 curated edge-knowledge test cases.

Use this module to teach the team how to define a skill, write its evaluation, and reason about model differences before any infrastructure is involved.

2. OpenClaw_AgentHarness/ — Multi-Agent Sandbox Pipeline

A Docker-Compose stack that introduces OpenClaw as the sandbox / gateway. Three least-privilege Agents collaborate in a shared workspace through a self-closing loop:

Only Agent C holds the exec capability, and it is constrained via tools.exec.allowedPaths to a single subdirectory. Models are served through GitHub Copilot’s built-in provider (Claude Opus 4.7).