Feature: Multi-LLM Support Implementation

[misran3]

Multi-LLM Support Implementation

Summary

This PR adds comprehensive multi-LLM provider support to the Android Use agent, enabling users to choose between OpenAI, Anthropic (Claude), Google Gemini, and AWS Bedrock. The implementation uses Pydantic AI for a unified interface and structured outputs, making the codebase provider-agnostic while improving type safety and reliability.

Key Features

✅ Multi-Provider Support

• OpenAI (GPT-4o, GPT-4o-mini, etc.)
• Anthropic Claude (Claude Sonnet 4, etc.)
• Google Gemini (Gemini 2.0 Flash, etc.)
• AWS Bedrock (Claude on Bedrock)

✅ Structured Outputs

• Pydantic models for type-safe action validation
• Automatic schema validation
• Better error messages

✅ Environment-Based Configuration

• Simple provider switching via LLM_PROVIDER env var
• Provider-specific API key management
• Optional model overrides per provider

✅ Comprehensive Testing

• 11/11 unit tests passing
• Provider initialization tests
• Error handling tests
• Integration tests

Architecture Changes

Before

# Direct OpenAI client
client = OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))
response = client.chat.completions.create(model="gpt-4o", ...)
decision = json.loads(response.choices[0].message.content)

After

# Unified LLM manager with structured outputs
llm_manager = LLMManager()  # Reads LLM_PROVIDER env var
action = await llm_manager.get_decision(goal, screen_context)
# Returns validated Pydantic model (TapAction, TypeAction, etc.)

New Files

• action_models.py - Pydantic models for structured action outputs
• llm_manager.py - Multi-provider LLM manager
• tests/test_action_models.py - Action model validation tests
• tests/test_llm_manager.py - LLM manager initialization tests
• tests/test_kernel_integration.py - Integration tests
• tests/MANUAL_TEST_RESULTS.md - Manual testing template
• examples/openai_example.sh - OpenAI usage example
• examples/anthropic_example.sh - Anthropic usage example
• examples/gemini_example.sh - Gemini usage example
• examples/bedrock_example.sh - Bedrock usage example

Modified Files

• kernel.py - Refactored to use LLM manager, async/await pattern
• requirements.txt - Updated to pydantic-ai-slim[openai,anthropic,google,bedrock]
• README.md - Added multi-LLM configuration instructions

Usage

Quick Start

# Choose your provider
export LLM_PROVIDER="openai"  # or anthropic, gemini, bedrock

# Set API key
export OPENAI_API_KEY="sk-..."

# Optional: Override default model
export OPENAI_MODEL="gpt-4o"

# Run the agent
python kernel.py

Provider-Specific Examples

# OpenAI
export LLM_PROVIDER="openai"
export OPENAI_API_KEY="sk-..."

# Anthropic Claude
export LLM_PROVIDER="anthropic"
export ANTHROPIC_API_KEY="sk-..."

# Google Gemini (cheapest option)
export LLM_PROVIDER="gemini"
export GOOGLE_API_KEY="..."

# AWS Bedrock
export LLM_PROVIDER="bedrock"
export AWS_PROFILE="default"

Available Models

Provider	Default Model	Override Env Var
OpenAI	gpt-4o	OPENAI_MODEL
Anthropic	claude-sonnet-4	ANTHROPIC_MODEL
Gemini	gemini-2.0-flash-exp	GEMINI_MODEL
Bedrock	anthropic.claude-sonnet-4-20250514-v1:0	BEDROCK_MODEL

Testing

Automated Tests

pytest tests/ -v
# 11/11 tests passing

Manual Testing

• Error handling verified via unit tests
• Real API testing requires valid credentials
• See tests/MANUAL_TEST_RESULTS.md for testing checklist

Breaking Changes

⚠️ Migration Required

1. Environment Variables: Must set LLM_PROVIDER env var

   # Before
   export OPENAI_API_KEY="sk-..."
   
   # After
   export LLM_PROVIDER="openai"
   export OPENAI_API_KEY="sk-..."

2. Async Pattern: run_agent() is now async

   # Before
   run_agent(goal)
   
   # After
   asyncio.run(run_agent(goal))

3. Dependencies: New requirement format

   # Before: openai>=1.12.0
   # After: pydantic-ai-slim[openai,anthropic,google,bedrock]
   

Benefits

1. Cost Optimization: Switch to cheaper providers (Gemini) for non-critical workflows
2. Reliability: Fallback to alternative providers if one is down
3. Type Safety: Pydantic models catch errors before execution
4. Developer Experience: Clear error messages, better testing
5. Future-Proof: Easy to add new providers (Ollama, Llama, etc.)

Next Steps

• Manual testing with real API keys
• Refactor package structure for modularity and PyPI distribution
• Add Ollama/Llama support for local models
• Optional retry/fallback logic between providers

---

Questions & Contact

Have questions about this PR or want to discuss implementation details?

📧 Email: mohammed.misran@pitt.edu

📅 Schedule a call: Calendly

Feel free to reach out for clarifications, feedback, or collaboration opportunities!

[misran3]

Hey @ethanjlimgit, I've raised this as a draft PR to be transparent about the work in progress and get early feedback.

Current Status:

• All unit tests passing (11/11)
• Currently performing manual testing with real API keys
• Verifying each provider works as expected

Feel free to review the code, ask questions, or share thoughts! I'm happy to discuss any aspect of the implementation or make adjustments based on your feedback.

Thanks!

[misran3]

Hey @ethanjlimgit, I've raised this as a draft PR to be transparent about the work in progress and get early feedback.

Current Status:

* All unit tests passing (11/11)

* Currently performing manual testing with real API keys

* Verifying each provider works as expected

Feel free to review the code, ask questions, or share thoughts! I'm happy to discuss any aspect of the implementation or make adjustments based on your feedback.

Thanks!

Hey @ethanjlimgit, wondering if you got a chance to review this PR?