Skip to content

Latest commit

 

History

History
133 lines (97 loc) · 5.03 KB

agents.yml.md

File metadata and controls

133 lines (97 loc) · 5.03 KB
Raw

agents.yml Usage Guide

The agents.yml configuration file is inspired by AutoGen and is compatible with AIConfig, with the additional condition that the description field must be provided as it is used by GitHub Copilot Chat and LLM-based routing.

When this extension first starts up, it will look in the following locations for an agents.yml file in the your home directory and in ~/.agents/. If it does not find one it will create a new default config file.

You can ask to see your config file at any time:

@agents /config

Any changes to this file will be immediately applied.

This extension will also scan for and read any AIConfig files (ending in aiconfig.json) in ~/ and ~/.agents/.

Structure of agents.yml

config_list

This section is used to define the models and their corresponding API keys. For example:

config_list:
  - model: gpt-4
    api_key: YOUR_API_KEY
  - model: gpt-3.5-turbo
    api_key: YOUR_API_KEY

This section is optional if you have a OAI_CONFIG_LIST environment variable or JSON file in ~/ or ~/.agents/ which could be shared by AutoGen and other libraries/frameworks.

agents

Copilot Agents uses some of the optional properties of the agent descriptors to determine which type of Agent to create:

  • agent.routing.team creates a RouterAgent - see Routing below.
  • agent.remote creates an AgentProtocolAgent - refer agentprotocol.ai
  • agent.cli.command creates a CliClientAgent - executes an Agent from a command line (useful for calling Python agents from Node.JS etc)
  • agent.llm_config.config_list - If an LLM has been configured the agent will use OpenAI, local LLMs etc
  • If No LLM has been configured, will attempt to use GitHub Copilot's chat completion service.

If agent.vscode === true

Agent Properties

Each agent can have the following properties:

  • name: The name of the agent. /[a-zA-Z_]+/ - avoid dots, slashes etc.

  • description: This will appear after the agent name when you mention it by @name.

  • schema_version: Required, only if you want to use with AIConfig

  • metadata: Skip this unless you're using with AIConfig

  • prompts: Named prompts which your model can use. Some of the agents provided by this extension attempt to use a prompt named background.

  • vscode: Set to true to have the Agent available by mentioning it by @{name} in VS Github Copilot Chat.

  • icon: One of the VS Code icon codes or a URL to an online image.

  • models: A list of model names to filter - eg ['gpt-4', 'gpt-3.5-turbo'] These reference models defined in agents[i].metadata.models, config_list and OAI_CONFIG_LIST.

  • sample_request: When the user clicks this agent in /help, this text will be submitted to this slash command.

  • commands: A list of { name, description } and optionally sampleRequest. The agent needs to be coded to process these.

The metadata/models section can be used to give aliases to the models, eg:

agents:
  - name: 10xDeveloper
    description: An experienced software developer
    metadata:
      default_model: default
      models:
        default:
          model: gpt-4
          temperature: 0.5
        fast:
          model: gpt-3.5-turbo
          temperature: 0.1
        fun:
          model: gpt-3.5-turbo
          temperature: 2.0

Agent Protocol

You can connect to an Agent running locally or remotely using agents[i]/remote.

Example:

 - name: AutoGPT
    description: The power of AI accessible to everyone
    icon: https://avatars.githubusercontent.com/u/130738209?s=32
    vscode: true
    remote:
      base_url: http://localhost:8000

Agent as CLI Process

You can spawn a new process to run an Agent written in any language.

- name: evo_ninja
  description: Research, analyse and build software
  icon: https://evo.ninja/favicon.ico
  vscode: true
  git:
    repo: https://github.com/polywrap/evo.ninja
    branch: dev
    baseDir: ~/.agents/evo.ninja
  cli:
    command: cmd.exe
    args: ["/K", "yarn", "start"]
    cwd: ~/.agents/evo.ninja
    wait_for: "Enter another goal: "

The example above downloads and installs the agent from source, pulling the latest changes on every launch.

Routing

All Agents automatically register themselves with the AgentRegistry upon construction.

Note: Name your agents carefully - if an Agent attempts to send a message to another Agent by role (eg "developer") and an Agent exists with a name matching that role, it will receive the message.

The RouterAgent and WorkflowManager use context fields to determine which Agent to route each user message to.

For example, a request to "create a new react web app" may result in the following context fields being set:

  • role: "developer", "tester" etc
  • language: ["javascript", "typescript"]
  • platform: ["web"]
  • framework: ["react"]