v0.1-alpha - Multi-step Agents & Planning

[kreneskyp]

Description

This PR is a major rewrite of AgentProcess to support multi-step agents and advanced workflows. This is part of a roadmap to provide incredibly flexible LLM agents at scale.

This branch is still an alpha release but is available for testing.

Langchain

The AgentProcess has been converted to use langchain chains, LLM, and memory. This is introduces support for multi-step LLM processes.

The default agent is now a zero-shot agent that runs a root chain. A LLMToolChooser may be used to replicate decision making of an agent. Future releases will introduce additional agent types.

Chain Visualizer

A chain visualizer was added to help understand how chains are structured.

Planning Chain

The demo planning chain creates plans for user requests and then executes them when authorized.

Chat replacing Tasks

The Chat view is now known as Chat. Tasks still exist as chains that run the chat and actions taken from the chat.

v0.2 Roadmap

• More visualizer features and editing mode
• Support for langchain tools
• Additional Agent types

Changes

[List out the changes you've made in this pull request. Be as specific as possible.]

How Tested

[Explain how you tested this pull request. Include any relevant steps or scripts.]

TODOs

• index still defaults to new task view
• New Chat button creates a new chat but doesn't clear the existing chat. Workaround: refresh page.
• agent config doesn't support selecting a chain
• deprecate agent tool config

For follow up:

• add support for adding memory to chains.
  • Artifact Memory #41
  • Chat message memory #42
• Tasks & Artifacts do not refresh  #40 Task/Artifact list doesn't update without a refresh.
• THINK/THOUGHT langchain refactor #39
  • THOUGHT messages do not include usage
  • THINK/THOUGHT messages is currently hooked to the chain and not LLM calls.
• Planner plans do not execute automatically when in autonomous mode