tape_improver

Tape Improver

This example show an Agentic Tape Improver that use a pre-generated tape from the Data Science Agent and generate better step. The CodeImprover agent is composed of 3 other agents: AgentSelector, StepSelector and StepRewriter.

Setup

The CodeExecutor use Podman for container execution. You must install it on your machine (see instructions).

In some cases you will have to set DOCKER_HOST environment variable to make Podman accesible for TapeAgents code, e.g. DOCKER_HOST=http+unix:///var/run/docker.sock. See the output of podman machine start for the path to the socket.

Execution

1. In a terminal, run the first app that load a pre-saved bad tape

python -m examples.tape_improver.tape_improver agent

2. Under the dropdown Run a transform, select improve_code. This run the CodeImprover multi-agent which update the tape.

3. In a separated terminal, run the second app:

python -m examples.tape_improver.tape_improver improver

4. Copy the field author_tape_id from the Tape's Metadata of the first app and paste it in the second app under Load tape by id. In the second app, you can now see the tape produced by the CodeImprover agent to improve the initial tape.

5. In the first app you can click on Run Loop to continue the tape from the improved step.

See outputs folder for the code files and images that the CodeExecutor generated.