functionsmith 0.1.0

Functionsmith is a general-purpose problem-solving agent using dynamic function calling

Functionsmith

Overview

Functionsmith is a general-purpose problem-solving agent using dynamic function calling.

USING THIS AGENT IS UNSAFE. It directly runs LLM-produced code, and thus should only be used for demonstration purposes.

Until there is a PyPi package, install with:

pip install "git+https://github.com/google/earthengine-community.git#egg=functionsmith&subdirectory=experimental/functionsmith"

To run with the default task investigating a CSV file with airport data:

export GOOGLE_API_KEY=my-api-key
wget https://raw.githubusercontent.com/davidmegginson/ourairports-data/refs/heads/main/airports.csv
functionsmith_cli

This agent uses dynamic function calling, which means that instead of relying on a fixed set of tools predefined in the agent in normal LLM function calling, we let the agent itself write with all the functions it needs.

The functionsmith system prompt asks the agent to first write any low-level function it needs, as well as tests for them. The agent loop will try to run these functions and ask the LLM to make corrections if necessary. Once all the functions are ready, the agent will write and run the code to solve the actual user task.

The agent does not use function calling features of LLM clients. Instead, it simply tries to parse all the python or tool_use sections present in the raw LLM output. It keeps all function definitions as well as their source code in memory. Each call to the LLM is preceded by the function definitions to let the LLM know what functions are available locally.

The functions are not saved permanently, though this feature can be added.

TODO(simonf): add a notebook version of the general-purpose agent, as well as an Earth Engine-specific notebook agent.

Attribution

Functionsmith was written by Simon Ilyushchenko ([e-mail]). I am grateful to Renee Johnston and other Googlers for implementation advice, as well as to Earth Engine expert advisors Jeffrey Cardille, Erin Trochim, Morgan Crowley, and Samapriya Roy, who helped me choose the right training tasks.