initial commit

nalbion · Feb 13, 2024 · a91f41c · a91f41c
1 parent 287aefb
commit a91f41c
Show file tree

Hide file tree

Showing 5 changed files with 298 additions and 1 deletion.
diff --git a/OAI_CONFIG_LIST.md b/OAI_CONFIG_LIST.md
@@ -0,0 +1,26 @@
+# OAI_CONFIG_LIST
+
+The `OAI_CONFIG_LIST.json` file is shared with [AutoGen] and allows you to keep your API keys separate from the `agent.yml` file which you may want to share with others.
+
+You can also define `base_url`, `api_type`, `api_version` for use with Open Source models.
+
+You can get your Open AI API key from the [API keys](https://platform.openai.com/api-keys) section in the platform site.
+
+The `model` names here are sent to the Chat Completion service, so must match with the model names they use.
+
+Example:
+
+```json
+[
+  {
+    "model": "gpt-4",
+    "api_key": "sk-1234"
+  },
+  {
+    "model": "gpt-4-1106-preview",
+    "api_key": "sk-7890"
+  }
+]
+```
+
+[AutoGen]: https://github.com/microsoft/autogen
diff --git a/README.md b/README.md
@@ -1,5 +1,9 @@
 # agent-adapters
 
+Framework for configurable AI Agents.
+
+Inspired by with [AutoGen](https://github.com/microsoft/autogen) and compatible with [AIConfig](https://aiconfig.lastmileai.dev/) and [Agent Protocol](https://agentprotocol.ai/). Can also download invoke and invoke arbitrary Agents as a subprocess.
+
 ## Installation
 
 To install this library, you can use npm:
@@ -13,9 +17,32 @@ npm install agent-adapters
 Import the library in your TypeScript/JavaScript file:
 
 ```typescript
-import * as agentAdapters from "agent-adapters";
+import { readAgentConfig, createAgent } from "agent-adapters";
 ```
 
+## agents.yml
+
+The `agents.yml` file is inspired by [AutoGen](https://github.com/microsoft/autogen) and is compatible with [AIConfig](https://aiconfig.lastmileai.dev/), with the additional condition that the `description` field must be provided as it is used by GitHub Copilot Chat and LLM-based routing.
+
+`createAgent()` uses some of the optional properties of the `AgentConfig` to determine which type of Agent to create:
+
+- `agent.routing.team` -> `RouterAgent` - see **Routing** below.
+- `agent.remote` -> `AgentProtocolAgent` - refer [agentprotocol.ai](https://agentprotocol.ai/)
+- `agent.cli.command` -> `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 will use OpenAI, local LLMs etc
+- If No LLM has been configured, will attempt to use GitHub Copilot
+
+If `agent.vscode === true` the Agent will be available by mentioning it by `@{name}` in VS Github Copilot Chat (when [ChatAgents2 proposal](https://code.visualstudio.com/api/advanced-topics/using-proposed-api#using-a-proposed-api) is released).
+
+### Routing
+
+All Agents automatically register themselves with the `AgentRegistry` upon construction.
+Agents may use `AgentRegistry.searchAgents()` to search by name or `role`, and matching Agents will be ranked based on the `context`.
+
+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.
+
+- role: "developer", "tester" etc
+- context: `{ language: ["javascript", "typescript"], platform: ["web"], framework: ["react"], ... }`
 
 ## Development
 

diff --git a/agents.yml.md b/agents.yml.md
@@ -0,0 +1,133 @@
+# 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](https://platform.openai.com/api-keys). For example:
+
+```yaml
+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](https://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](https://code.visualstudio.com/api/references/icons-in-labels#icon-listing) 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:
+
+```yaml
+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:
+
+```yaml
+ - 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.
+
+```yaml
+  - 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"]
+
+[AIConfig]: https://aiconfig.lastmileai.dev/
+[AutoGen]: https://github.com/microsoft/autogen
diff --git a/examples/default/agents.yml b/examples/default/agents.yml
@@ -0,0 +1,30 @@
+config_list:
+  - model: gpt-4
+    api_key: YOUR_API_KEY
+  - model: gpt-3.5-turbo
+    api_key: YOUR_API_KEY
+
+agents:
+  - name: my-agent
+    description: Uses OpenAI's GPT-4 LLM
+    vscode: true
+    models: ["gpt-4"]
+    prompts:
+      - name: background
+        input: You are a helpful assistant
+    routing:
+      roles:
+        - developer
+
+  - name: evo_ninja
+    description: Research, analyse and build software
+    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: "
diff --git a/examples/full/agents.yml b/examples/full/agents.yml
@@ -0,0 +1,81 @@
+config_list:
+  - model: gpt-4
+    api_key: YOUR_API_KEY
+  - model: gpt-3.5-turbo
+    api_key: YOUR_API_KEY
+
+agents:
+  - name: 10xDeveloper
+    description: An experienced software developer
+    # /agents/[]/metadata/models is an AIConfig convention (https://aiconfig.lastmileai.dev)
+    metadata:
+      default_model: default
+      models:
+        default:
+          model: gpt-4
+          temperature: 0.5
+        fast:
+          model: gpt-3.5-turbo
+          temperature: 0.3
+
+    # /agents/[]/models can be used to filter the OAI_CONFIG_LIST model configs in /config_list
+    models: ["gpt-4"]
+
+    # /agents/[]/llm_config is used at runtime, but if not provided can be generated on initialisation
+    # from /config_list, /agents/[]/models and/or /agents/[]/metadata/models
+    llm_config:
+      config_list:
+        - model: gpt-4
+          api_key: YOUR_API_KEY
+      model_settings:
+        default:
+          model: gpt-4
+          temperature: 0.5
+
+    # Should the agent be available in the VSCode extension
+    vscode: true
+    # Fields used by the RouterAgent to determine which agent to use
+    routing:
+      roles: ["developer"]
+      context:
+        languages: ["c++", "python", "javascript", "java", "c#"]
+    prompts:
+      - name: "background"
+        input: |
+          You are an experienced software engineer who practices TDD and CI/CD.
+          You are familiar with the following:
+           - languages: C++, Python, JavaScript, Java, and C#.
+           - frameworks: React, Angular, Vue, Django, Flask, Spring, and .NET.
+           - databases: MySQL, PostgreSQL, MongoDB, and Redis.
+           - cloud platforms: AWS, Azure, and GCP.
+           - CI/CD tools: Jenkins, CircleCI, and TravisCI.
+           - containerization tools: Docker and Kubernetes.
+
+  - name: router
+    description: A router agent
+    vscode: true
+    models: ["gpt-4"]
+    routing:
+      team: ["10xDeveloper", "python_demo"]
+
+  - name: scrum
+    description: My AI scrum team
+    vscode: true
+    workflow_base_path: ~/.agents/workflows/scrum
+
+  - name: command_line
+    description: Talk to the command line
+    vscode: true
+    cli:
+      command: cmd.exe
+      args: ["/K"] # /K keeps the process open
+      # args: ["/C"] # /C closes the process after execution
+      # daemon: true
+
+  - name: python_demo
+    description: Python CLI client demo
+    vscode: true
+    cli:
+      # command: %HOME_PATH%\.pyenv\pyenv-win\shims\python
+      command: python
+      args: ["../AutoGen_Agent_Demo/agents/demo.py"]