AI工作流开源项目

Example manifest

apiVersion: adl.inference-gateway.com/v1
kind: Agent
metadata:
  name: customer-support
  description: AI agent for handling customer inquiries
  version: "1.0.0"
spec:
  capabilities:
    streaming: true
    pushNotifications: true
    stateTransitionHistory: true
  agent:
    provider: deepseek
    model: deepseek-v4-flash
    systemPrompt: |
      You are a professional customer support agent.
    maxTokens: 4096
    temperature: 0.3
  tools:
    - id: knowledge_search
      name: knowledge_search
      description: Search the company knowledge base
      tags:
        - knowledge
        - search
      schema:
        type: object
        properties:
          query:
            type: string
            description: The search query
        required:
          - query
  skills:
    - id: incident-response
      bare: true
      name: incident-response
      description: How to triage a paged production incident, draft an initial response, and notify stakeholders.
      license: Apache-2.0
      tags:
        - operations
        - incident
  server:
    port: 8080
    debug: false
  language:
    go:
      module: github.com/company/customer-support-agent
      version: "1.26"
  development:
    sandbox:
      flox:
        enabled: true
      devcontainer:
        enabled: false
    ai:
      enabled: true

Tools vs Skills

ADL distinguishes two ways for an agent to act:

• Tools (spec.tools[]) are function-call entrypoints. Each tool has a JSON Schema for its inputs and is generated as code in the target language. Use a tool when the agent needs to invoke a deterministic operation (query a database, send an email, call an API).
• Skills (spec.skills[]) are markdown playbooks injected into the agent's system prompt at startup. They're either pulled from the skills registry or scaffolded blank with bare: true. Use a skill when you want to teach the agent a workflow, policy, or response pattern in natural language.