Xingyao Wang 537fb7d985
feat: convert agent prompts into structured Jinja2 templates (#3360)
* commit jinja draft

* remove extra file

* update system prompt

* remove github message

* update prompts

* add prompt manager and its tests

* use prompt manager for codeact and bump version

* fix integration tests

* fix lint

* simplify test case

* update system

* fix integration tests

* update credit path for aider

* Update CREDITS.md

Co-authored-by: Engel Nyst <enyst@users.noreply.github.com>

---------

Co-authored-by: Engel Nyst <enyst@users.noreply.github.com>
2024-08-18 16:38:46 +00:00
..

CodeAct Agent Framework

This folder implements the CodeAct idea (paper, tweet) that consolidates LLM agents actions into a unified code action space for both simplicity and performance (see paper for more details).

The conceptual idea is illustrated below. At each turn, the agent can:

  1. Converse: Communicate with humans in natural language to ask for clarification, confirmation, etc.
  2. CodeAct: Choose to perform the task by executing code
    • Execute any valid Linux bash command
    • Execute any valid Python code with an interactive Python interpreter. This is simulated through bash command, see plugin system below for more details.

image

Plugin System

To make the CodeAct agent more powerful with only access to bash action space, CodeAct agent leverages OpenDevin's plugin system:

Demo

https://github.com/OpenDevin/OpenDevin/assets/38853559/f592a192-e86c-4f48-ad31-d69282d5f6ac

Example of CodeActAgent with gpt-4-turbo-2024-04-09 performing a data science task (linear regression)

Work-in-progress & Next step

[] Support web-browsing [] Complete the workflow for CodeAct agent to submit Github PRs