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).

This module teaches tool allowlists, workspace isolation, gateway-based credential rotation, and self-correcting loops.

3. CodingAgent/ — Context Optimization & Error Correction

A more advanced OpenClaw pipeline focused on two production techniques:

1. Context Optimization — the coder Agent only loads Skill documents that the spec explicitly references via @FILE.md syntax (workspace/skills/PYTHON_STYLE.md, ALGO_PATTERNS.md, ERROR_HANDLING.md, TESTING.md). This keeps the prompt window lean and predictable.
2. Error Correction — a dedicated diagnoser Agent translates a captured Python traceback into a machine-readable Patch Plan (DIAGNOSIS.md), which the next iteration of the coder consumes.

Three roles — coder, runner, diagnoser — each with a strict tool allowlist, demonstrate how to scale a self-healing coding loop while still controlling token spend.

4. AgentHarness_HostedAgent/ — Hosted Agent on Microsoft Foundry

The flagship module. It re-implements the Skill_Testing_Agent logic on top of a managed-style hosted-agent harness with the four-layer architecture:

       ┌──────────────────────────┐    execute(name,input) -> str       ┌──────────┐
       │   Orchestrator (Brain)   │ ───────────────────────────────────▶│  Hands   │
       │  Foundry hosted agent    │                                     │ (sandbox │
       │  (default: gpt-5.5)      │  emit_note / get_events / list_*    │  cattle) │
       └──────────────┬───────────┘                                     └──────────┘
                      │
                      ▼
       ┌──────────────────────────┐
       │       SessionStore       │  append-only log, decoupled from the brain context window
       └──────────────────────────┘

• Brain — A SkillTestingHarness agent running on ResponsesHostServer (Foundry hosted agent / Responses protocol). Exposes only 5 tools: execute, list_tools, list_models, get_events, emit_note.
• Hands — A SandboxPool of disposable, stateless workers. Seven hands are registered: list_test_cases, run_business_agent, validate_format, judge_rubric, craft_attack, next_attack_prompt, run_full_benchmark.
• Session — A file-level append-only log, decoupled from the brain’s context window. Replayable via wake(session_id).
• Vault — Credentials registered by logical name; sandboxed hands never see raw tokens.

The agents under test are two Microsoft Foundry deployments: DeepSeek-V4-Flash and gpt-5.5. The module ships a CLI runner (main_local.py) for fast model comparison and a hosted entrypoint (main.py) that speaks the Responses protocol.

This module teaches the Brain / Hands / Session / Vault separation and how to keep an enterprise harness debuggable, restartable, and credential-safe.

5. skill-testing-harness-deploy/ — Deploy to Microsoft Foundry with azd

An azure.yaml + Bicep package that deploys the harness as a hosted agent (host: azure.ai.agent) using azd up. Demonstrates the full IaC story: container resources, startup command, remote build, and the azure.ai.agents extension.

6. ppt/ — Workshop Slide Deck

The supporting slides delivered alongside the code, in suggested presentation order:

1. 00.enterprise-agent-evolution.pptx — Why enterprise Agents, and how the field has evolved.
2. 01.openclaw-orchestration-deep-practice.pptx — Deep-dive on OpenClaw multi-agent orchestration.
3. 02.token-economics-cost-control.pptx — Token economics and cost control under real workloads.
4. 03.skill-definition-and-testing-standards.pptx — Skill definition and adversarial-testing standards.
5. 04.coding-agent-token-report.pptx — A token-usage post-mortem for the coding-agent loop.
6. 05.agent-harness-.pptx — Walkthrough of the Brain / Hands / Session / Vault harness.

---

Suggested Learning Path (one-day agenda)

A workable rhythm for a single day (≈ 6–8 hours):

---

Prerequisites

The whole workshop runs against the same baseline environment:

1. Python 3.12+
2. GitHub Copilot subscription (with seats enabled for the attendees’ GitHub accounts).
3. GitHub Copilot CLI installed and signed in:

   # Install GitHub CLI first if you don't have it: https://cli.github.com/
   gh extension install github/gh-copilot
   # Or use the dedicated copilot CLI per https://github.com/github/copilot-cli
   copilot auth login
   

4. Microsoft Foundry project — an Azure subscription with a Foundry project endpoint, plus model deployments for DeepSeek-V4-Flash and gpt-5.5 (names are configurable via env vars).

Additional per-module needs:

---

Audience

Internal engineering teams who already ship software in production and now need to:

• Define skills with testable contracts instead of free-form prompts.
• Stand up multi-agent pipelines that respect the principle of least privilege.
• Reason about token economics and context windows under real workloads.
• Migrate from a laptop prototype to a Foundry-hosted, IaC-deployed agent runtime.

The materials are intentionally code-first. Every concept introduced in the slide deck has a runnable counterpart in one of the directories above.

中文版 README 见 README.zh.md。