Exploring agents-cli and alternatives (vyomi, localgcp):
- To test and spin up some GCP compliant Gemini Enterprise Agent Platform environments without spending a lot of time, money.
- To test various TensorFlow configurations and deployments (
logistic,linear,forecasting, batch sizes, distributed gpu, shardedTFRecords, etc.) that come up in the GCP PMLE. - Opted for the official Google
agents-clisince multi-Agent examples can be entirely Containerized, used locally, migrated to an actual GCP environment, and since it's likely used in various production enterprise GCP settings. - To review test patterns for GCP EAP Agent Models.
This is an ai-assisted-development effort using GitHub Copilot to quickly scaffold an agents-cli lab with a few kinds of tf examples (logistic, linear, forecasting).
The inbuilt agents-cli playground supports both local development/testing and publishing to GCP Gemini.
There are some limitations: I've chosen a configuration that's CPU only on a Mac.
- For ARM GPU support on newer Macs, one has to use
tensorflow-metalbut it only provides a single GPU backend. It doesn't offer much advantage for this limited demo. - I will probably tinker around with some kind of NVIDIA GPU with multiple Linux Containers in the future.
- I had GitHub Copilot expand and add to the available Agent Tooling:
run_experiment,run_tests.
Example batch sizes, distributed gpu, multiworker, sharded TFRecords, memory.pin configs can be found in train.yaml.
Train Models:
docker compose run train --demo forecast --set data.num_shards=16
docker compose run train --demo logistic --set data.num_shards=16
docker compose run train --demo linear --set data.num_shards=16
# Multiple works with global batch
docker compose --profile multiworker up --abort-on-container-exitSpin up the Agent:
- Get some Google AI Studio credentials and configure .env.
- This is probably the fastest, easiest, cheapest way to do this.
docker compose --profile agent up- http://localhost:8080/dev-ui/
- The Agent integrates with GCP Gemini 3.8 in the cloud and handles LLM interactions against tools defined here.
- Screenshots are supplied in _lab.