Coding Agent on Agent-Sandbox and LangGraph

This guide documents the complete setup and deployment of a coding agent that runs locally on Kubernetes using Kind (Kubernetes in Docker) with the agent-sandbox controller. The agent uses locally-cached transformer models to generate and execute Python code with automatic error correction.

Architecture

Components

  1. Agent-Sandbox Controller: Kubernetes operator that manages sandboxed workloads
  2. Init Container: Downloads and caches the ML model on first run
  3. Main Agent Container: Runs the LangGraph-based coding agent with local model inference
  4. Persistent Volume Claim: Stores the cached model

Technology Stack

  • Kubernetes: Kind (Kubernetes in Docker)
  • ML Framework: Transformers, PyTorch
  • Agent Framework: LangGraph
  • Model: Salesforce/codegen-350M-mono (350M parameters)
  • Language: Python 3.13

Prerequisites

Required Software 

# Verify installations
go version              # Go 1.19+
make --version          # GNU Make
python3 --version       # Python 3.13+
docker --version        # Docker 20.10+
kind --version          # Kind 0.17+
kubectl version --client # kubectl 1.25+

Setup Instructions

1. Deploy Agent-Sandbox to Kind

You can find agent-sandbox setup instructions here.

# Verify installation
kubectl -n agent-sandbox-system wait --for=condition=Ready pod -l app=agent-sandbox-controller

The kubectl wait command will exit when the pod is ready.

2. Set Up HuggingFace Token 

# Export your HuggingFace token
export HF_TOKEN='your_huggingface_token_here'

Get a token from HuggingFace.

3. Navigate to the examples, coding-agent directory 

cd examples/coding-agent

4. Set your huggingface token

Update <HF_TOKEN> placeholder in deployment.yaml:

apiVersion: v1
kind: Secret
metadata:
  name: coding-agent-hf-token
  namespace: default
type: Opaque
stringData:
  token: "<HF_TOKEN>"  # Replace with your actual HuggingFace token

5. Build and Deploy 

# Build init container image
docker build -f Dockerfile.init -t model-downloader:latest .

# Build agent image
docker build -t coding-agent:latest .

# Load images into Kind
kind load docker-image model-downloader:latest --name agent-sandbox
kind load docker-image coding-agent:latest --name agent-sandbox

# Deploy to Kubernetes
kubectl apply -f deployment.yaml

# Watch init container download model (first run only, takes 3-5 minutes)
kubectl logs -f coding-agent-sandbox -c model-downloader

# Wait for pod to be ready
kubectl wait --for=condition=ready pod -l app=coding-agent --timeout=600s

5. Use the Agent 

# Attach to the agent & hit enter to get interraction prompt
kubectl attach -it coding-agent-sandbox

# Example usage:
# You: Generate code to calculate factorial of 5
# [Agent generates, executes, and potentially fixes code automatically]

Workflow

Agent Execution Flow

  1. User Input: User provides a coding task
  2. Code Generation: LLM generates Python code based on the task
  3. Execution: Code is executed in a sandboxed environment
  4. Error Detection: If execution fails, error is captured
  5. Auto-Fix: LLM attempts to fix the code (up to 3 iterations)
  6. Result: Final code and execution output are displayed

Troubleshooting

Out of Memory Errors

Problem: Pod crashes with OOM errors.

Solution: Increase memory limits in deployment.yaml:

limits:
  memory: "24Gi"  # Increase from 16Gi

Slow Code Generation

Problem: Takes 30-60 seconds to generate code. Expected: The 350M model runs on CPU without GPU acceleration. This is normal for local inference. Options:

  • Accept the performance
  • Use a smaller model
  • Add GPU support to your Kind cluster
  • Use a hosted API instead

Switching Models

To use a different model, update these files:

download_model.py 

MODEL_ID = "<your-model-id>"  # e.g., "bigcode/starcoder2-3b"

coding_agent.py 

def __init__(self, model_id: str = "<your-model-id>", hf_token: str = None):

After changing models:

# Delete PVC to remove old model
kubectl delete pvc models-cache-pvc
# Rebuild and redeploy
docker build -f Dockerfile.init -t model-downloader:latest .
docker build -t coding-agent:latest .
kind load docker-image model-downloader:latest --name agent-sandbox
kind load docker-image coding-agent:latest --name agent-sandbox
kubectl delete sandbox coding-agent-sandbox
kubectl apply -f deployment.yaml

Cleanup

Remove Agent Only 

kubectl delete sandbox coding-agent-sandbox
kubectl delete pvc models-cache-pvc
kubectl delete secret coding-agent-hf-token

Remove Agent-Sandbox Controller 

kubectl delete namespace agent-sandbox-system
kubectl delete crd sandboxes.agents.x-k8s.io

Remove Kind Cluster 

kind delete cluster --name agent-sandbox

References

Support

For issues:

  1. Check logs: kubectl logs coding-agent-sandbox -c agent
  2. Check init container: kubectl logs coding-agent-sandbox -c model-downloader
  3. Verify pod status: kubectl describe pod coding-agent-sandbox
  4. Review agent-sandbox controller: kubectl logs -n agent-sandbox-system -l app=agent-sandbox-controller

Last modified November 24, 2025: Docs: Add guides section (#140) (70a5a8f)