Demo Guide

Fastest demo: no external LLM

After installing the project:

python scripts/demo_incident.py

Expected story:

1. Ops Agent reads a simulated import-service timeout log.
2. Knowledge Agent retrieves the import operations manual.
3. Data Agent queries historical import_jobs.
4. Supervisor combines the three evidence sources.
5. The final answer explicitly says no repair action was executed.

For JSON output:

python scripts/demo_incident.py --json

Web demo

Local:

cp .env.example .env
docker compose up --build

Open:

http://127.0.0.1:8000/

Online demo:

https://agent.majhoon.site

Default demo credentials:

demo / demo

The Web Console is organized as a resource-oriented Enterprise SaaS console:

Home

BUILD
  Agents
  Knowledge
  Data sources

OPERATE
  Runs
  Evaluations
  Approvals

PLATFORM
  Integrations
  Credentials
  Policies

Settings

When AUTH_ENABLED=false, the application runs as development/admin. For shared demo environments, enable auth and configure both API tokens and the console login:

AUTH_ENABLED=true
AUTH_TOKENS={"user-token":"alice:user","operator-token":"operator:operator","approver-token":"reviewer:approver","admin-token":"admin:admin"}

CONSOLE_USERNAME=demo
CONSOLE_PASSWORD=demo
CONSOLE_ROLE=admin

The Web Console uses an HttpOnly session cookie. API clients can continue using Bearer tokens.

5-minute interview flow

1. Sign in and Home

Open the Web Console and sign in with demo / demo.

Show:

• Attention items
• Agent health
• Recent activity
• Runtime metrics

Emphasize that the home page answers "what needs attention?" rather than acting as a generic KPI dashboard.

2. Agents

Open the Agents directory.

Explain that agents are first-class resources, not hard-coded global sidebar tabs. Open Data Agent and show the common detail shell:

Agent
├── Playground
├── Runs
├── Evaluations
└── Configuration

Ask Data Agent:

统计每类数据的数量，并指出数量最多的类别。

Show:

• Schema-aware SQL generation
• SQL Guard
• query result
• chart
• Markdown report
• execution trace

3. Knowledge

Open Knowledge and add a document with tags and allowed roles.

Then open Knowledge Agent and ask a question to show:

• role-scoped retrieval
• citations
• retrieval trace

For a deeper RBAC demo, call the API with user/operator/approver Bearer tokens.

4. Human Approval

As an operator/admin, create a document delete or service restart approval.

Show:

pending
  ↓
approved
  ↓
consumed

Explain that the approval is bound to the exact Agent + Tool + Target and is single-use.

5. Supervisor

Open Supervisor and run the offline incident demo or ask:

为什么最近导入任务失败？请结合当前运维状态、知识库手册和历史数据给出排查结论。

Show the execution graph:

Incident
   ↓
Supervisor
   ↓
Ops / Knowledge / Data
   ↓
Synthesis

Then open the execution trace.

6. Runs

Open Runs and inspect the Supervisor run.

Show:

• status
• duration
• Request ID
• Correlation ID
• Trace
• Eval

This is the best place to explain the platform engineering layer.

7. Integrations / Credentials

Open:

• Credentials → model provider configuration
• Data sources → database resources
• Integrations → Prometheus / logs / Docker / Kubernetes / controlled services

Emphasize that the console models these as resources instead of exposing raw .env JSON.

What to emphasize

• LLM prompts are not the security boundary; Tools are.
• RBAC answers "who", Tool Policy answers "what capability", Approval answers "who authorized this risky target".
• Run/Trace/Eval make Agent behavior debuggable and regression-testable.
• The offline demo executes the actual Agent classes, not a mocked UI-only path.
• Supervisor only orchestrates child Agents and does not bypass their Tool boundaries.
• The product UI is resource-oriented: Agents, Runs, Approvals, Credentials, Data Sources, and Integrations are first-class objects.
• The platform intentionally keeps infrastructure small until measured scale requires more.