Autolang

Another take on BabyAGI, focused on workflows that complete. Powered by langchain.

Here's a simple demo: https://twitter.com/pictobit/status/1645504308874563584

Running

To run Autolang, follow these steps:

(Optional) Customize the tools provided to the agent

Install dependencies:

pip install -r requirements.txt

Copy the .env.example file to .env, then edit it:

cp .env.example .env

Run the script:

python -m autolang

Alternatively, run with Docker:

./run_docker.sh

Architecture

Autolang uses four main components:

Planner

Runs once at the start, it thinks of a strategy to solve the problem, and produces a task list.

Executor

A custom langchain agent, which implements ReAct to solve a single task in the plan. It can be provided any tools in the langchain format.

Learner

Here's the interesting part. The system holds an information context string, which starts empty. After each step, the learner merges the result with the current context, as a sort of medium-term memory

Reviewer

Assesses the current task list, based on the current completed tasks and generated info context, and reprioritizes the pending tasks accordingly

Next steps

Right now, the main limitation is the limited info context. As a next step, I'm planning on adding a "long term memory agent", that extracts information from the context, replacing it with a key. The executor agent will be provided a tool to retrieve these saved snippets if required.