AI agent that can create its own tools

Overview

The agent is currently setup to use a system prompt geared for AWS.

Prerequisites

• gnu compatible Make
• python 3.10+
• uv

Installation

The following can be used to install and update the dependencies.

make depsupdate

Environment

You must create a .env file with the need values.
The project default to OpenAI but can be easily changed to other providers as needed.
Depending on the features and providers you use fill in the vars as needed.

# AI Providers
OPENAI_API_KEY=
OPENAI_MODEL_NAME=gpt-4
ANTHROPIC_API_KEY=
GOOGLE_API_KEY=
GROQ_API_KEY=
PPLX_API_KEY=
HF_TOKEN=
# Search providers
SERPER_API_KEY=
TAVILY_API_KEY=
JINA_API_KEY=
GOOGLE_CSE_ID=
GOOGLE_CSE_API_KEY=

# Other ai tools
LANGFLOW_API_KEY=

### Tracing (optional)
LANGCHAIN_TRACING_V2=true
LANGCHAIN_ENDPOINT=https://api.smith.langchain.com
LANGCHAIN_API_KEY=
LANGCHAIN_PROJECT=auto_tool_agent

Usage

Argument help:

make app_help

If using for AWS you must assume the role you want to use prior to running the agent.

Dev mode uses uv run

uv run python -m auto_tool_agent "list all lambda functions in region us-east-1"

uv run python -m auto_tool_agent -f csv "list all s3 buckets in region us-east-1"

Installed run examples

Output CSV format to file data.csv using query of s3 buckets and their storage classes:

auto_tool_agent --format csv --output data.csv "list all s3 buckets and their storage class"

Output Markdown format to file data.md using query of top 5 hacker news articles:

auto_tool_agent.exe -f markdown -o data.markdown -i 5 -s generic.md "fetch me the top 5 articles from hacker news"

Output format examples

You can specify any of the supported output formats enhance the system prompt for better output formatting.

Markdown

uv run python -m auto_tool_agent -f markdown -o data.md 'list all s3 buckets in region us-east-1. Use a Markdown table with columns bucket_name, region'

CSV

uv run python -m auto_tool_agent -f csv -o data.csv 'list all s3 buckets in region us-east-1. Use CSV with fields bucket_name, region'

JSON

uv run python -m auto_tool_agent -f json -o data.json 'list all s3 buckets in region us-east-1. Use the following json schema [{"bucket_name": "string", "region": "string"}]' make run

Text

uv run python -m auto_tool_agent -f text -o data.txt 'list all s3 buckets in region us-east-1.'

Folders dev mode

• REPO_ROOT/src/auto_tool_agent/sandbox - sandbox folder where the agent will write its tools
• REPO_ROOT/src/auto_tool_agent/tools - tools available to agent that it cant edit
• REPO_ROOT/src/auto_tool_agent/system_prompts - system prompts folder
• REPO_ROOT/tools_archive - good tools that are currently disabled

Folders installed mode

• ~/.auto_tool_agent/sandbox - sandbox folder where the agent will write its tools
• ~/.auto_tool_agent/system_prompts - system prompts folder

Agent Logic

System prompt

The agent has access to system prompts in the src/auto_tool_agent/system_prompts folder.
You can change the default system prompt with the --system_prompt or -s param.
The default system prompt is aws.md which is tuned for AWS.

Tool folders

The agent has an always available set of tools to allow it to list, read and write files in the tools_tests folder.
The tools_tests folder is the only folder that the agent has access to.

Iterations

Each time the agent needs to create or update a tool it will start a new iteration. The maximum number of iterations defaults to 5 but can be changed with the --max_iterations or -m param.

Example 1:

• 1st iteration: Agent looks at user request and decides it needs to create a tool.
• 2nd iteration: Agent uses the tool to generate an answer.

Example 2:

• 1st iteration: Agent looks at user request and decides has all tools it needs and generates an answer.

Example 3:

• 1st iteration: Agent looks at user request and decides it needs to create a tool.
• 2nd iteration: Agent uses the tool but gets an error, looks looks at the broken tool and fixes it.
• 3rd iteration: Agent uses the tool to generate an answer.

Running in the Generic Development Container (GDC)

The GDC is a project I created years go to create a highly configurable IDE agnostic rapid development container.
Because the agent creates its own tools and those tools execute code locally where you are running the agent you may want some extra security / sandboxing. While using the GDC is not required to run the agent in a container it has many quality of life benefits for any developer.
Follow the GDC installation instructions then run the following command from the repository root:

run-dev-container.sh

Then in a separate terminal run:

docker exec -it auto_tool_agent-dev-1 bash -l

This will drop you at a prompt in the repository root inside the container.
Now you can run the agent in the normal way without worry that any code it generates causing any damage to your system.