Skip to content

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.4.0] - 2025-11-02

Added

  • Unified Tool/Function Calling: Provider-agnostic tool calling interface
  • ToolDefinition: Universal tool definition compatible with OpenAI, Anthropic, and Gemini
  • ToolParameter: JSON Schema-based parameter definitions
  • ToolCall: Unified format for tool calls from LLMs
  • ToolResult: Unified format for tool execution results
  • Same tool definitions work across all LLM providers

  • Single source of truth for tools - define once, use everywhere

  • OpenAI Tool Support (COMPLETE):

  • Updated OpenAIPrompt to accept tools and tool_choice parameters
  • Updated OpenAIMessage to support tool role and tool-related fields
  • Updated OpenAIResponse to include tool_calls
  • Automatic conversion between unified and OpenAI native formats
  • Multi-turn conversations with tool results
  • Agent implementation complete with tool call extraction

  • Example: examples/tool-calling-openai.py

  • Anthropic Tool Support (COMPLETE):

  • Updated AnthropicPrompt to accept tools and tool_choice parameters
  • Updated AnthropicMessage to support tool result content blocks
  • Updated AnthropicResponse to include tool_calls
  • Automatic conversion between unified and Anthropic tool_use format
  • Agent implementation complete with tool_use block parsing
  • Multi-turn conversations with tool results

  • Example: examples/tool-calling-anthropic.py

  • Gemini Tool Support (COMPLETE):

  • Updated GeminiPrompt to accept tools parameter
  • Updated GeminiResponse to include tool_calls
  • Automatic conversion between unified and Gemini function declaration format
  • Agent implementation complete with function_call parsing
  • Support for Gemini 1.5 Flash and Pro models
  • Example: examples/tool-calling-gemini.py

Changed

  • Made content field optional in all LLM response models (can be None when only tool calls)
  • All three LLM agents now support tool calling with consistent interface

Documentation

  • Added comprehensive tool calling examples for all three providers
  • Tool usage patterns and best practices
  • Multi-turn conversation examples with tool results
  • Examples show same tool definitions working across all providers

Testing

  • 17 unit tests for tool calling functionality (all passing)
  • Model instantiation tests
  • Format conversion tests
  • Validation tests

[0.3.3] - 2025-10-31

Fixed

  • LLM Integrations: Fixed on_complete() method signature in all LLM agents
  • Issue: Base class calls on_complete(output_data) with 1 argument, but LLM agents defined it with 2 parameters (input_data, output_data)
  • Error: TypeError: on_complete() missing 1 required positional argument: 'output_data'
  • Fix: Removed unused input_data parameter from all three LLM agents
  • Affected agents: OpenAI, Anthropic, Gemini

[0.3.2] - 2025-10-31

Fixed

  • LLM Integrations: Fixed emit_progress() method calls in all LLM agents (9 total fixes)
  • OpenAI Agent: Fixed 3 emit_progress() calls (calling_openai_api, openai_api_complete, openai_api_error)
  • Anthropic Agent: Fixed 3 emit_progress() calls (calling_anthropic_api, anthropic_api_complete, anthropic_api_error)
  • Gemini Agent: Fixed 3 emit_progress() calls (calling_gemini_api, gemini_api_complete, gemini_api_error)
  • Added missing await keyword for all emit_progress() calls
  • Fixed parameters to use named arguments: stage, progress, message, details
  • Added descriptive messages for better observability
  • Set appropriate progress values (0.0 for start, 1.0 for completion, 0.0 for errors)
  • Bug: emit_progress() was being called with only 2 arguments instead of required 4 arguments

[0.3.1] - 2025-10-31

Fixed

  • SwitchStep: Fixed emit_progress() method calls to use correct signature
  • Added missing await keyword for all emit_progress() calls
  • Fixed parameters to use named arguments: stage, progress, message, details
  • Added descriptive messages for better observability
  • Set appropriate progress values (0.0 for case selection, 0.5 for execution)

[0.3.0] - 2025-10-31

Added

  • SwitchStep: Multi-branch conditional routing for clean handling of 3+ cases
  • Switch/case pattern similar to programming languages
  • Selector function returns case key for routing
  • Optional default/fallback agent
  • Cleaner alternative to nested ConditionalStep blocks
  • Full event emission and observability

  • Example: examples/switch-step-routing.py

  • OpenAI Integration: Complete integration with GPT-4, GPT-4-turbo, and GPT-3.5-turbo models

  • OpenAIAgent for executing OpenAI API calls
  • Automatic token counting using tiktoken
  • Cost estimation per request
  • Support for system messages, conversation history, and JSON response format

  • Helper functions: create_simple_prompt(), create_system_prompt()

  • Anthropic Integration: Complete integration with Claude 3 models

  • AnthropicAgent for Claude Opus, Sonnet, and Haiku models
  • 200K token context window support
  • System prompt support (Claude's preferred format)
  • Cost estimation per request (per 1M tokens pricing)

  • Helper functions: create_simple_prompt(), create_system_prompt()

  • Google Gemini Integration: Complete integration with Gemini models

  • GeminiAgent for Gemini Pro, Gemini 1.5 Pro, and Gemini 1.5 Flash
  • Up to 1M token context window (Gemini 1.5)
  • Safety ratings extraction
  • Cost estimation per request
  • Configurable sampling parameters (temperature, top-k, top-p)

  • Helper function: create_simple_prompt()

  • 7 Comprehensive Examples:

  • basic-openai-agent.py - OpenAI integration basics
  • anthropic-agent.py - Anthropic Claude usage patterns
  • gemini-agent.py - Google Gemini demonstrations
  • multi-llm-fallback.py - Provider fallback strategies
  • llm-comparison.py - Side-by-side provider comparison
  • llm-cost-tracking.py - Cost tracking and budget management

  • conversation-agent.py - Multi-turn conversational agents

  • Optional Dependencies: Install integrations separately

  • pip install agent-orchestration-lib[openai]
  • pip install agent-orchestration-lib[anthropic]
  • pip install agent-orchestration-lib[gemini]
  • pip install agent-orchestration-lib[all-llm]

Features

  • Token counting and cost estimation for all LLM providers
  • Async/await support for all LLM API calls
  • Integration with existing retry strategies (including LLMFallbackRetry)
  • Event emission for observability
  • Type-safe Pydantic models for all inputs and outputs
  • Helper utilities for common prompt patterns

Documentation

  • Updated README with LLM integration examples
  • Added installation instructions for optional dependencies
  • Updated package docstring with LLM integration information

Changed

  • Version bumped to 0.3.0
  • Updated package description to highlight LLM integrations

[0.2.0] - 2025-01-24

Added

  • ConditionalStep for if/else branching logic in agent workflows
  • FlowAdapter to execute flows as agent blocks (nested flows)
  • LLMFallbackRetry strategy for automatic LLM model fallback on failures
  • Exported Flow class for multi-agent orchestration
  • Exported all retry strategies: ExponentialBackoffRetry, FixedDelayRetry, LinearBackoffRetry, NoRetry
  • Helper functions: retry_on_exception_type(), retry_on_error_message()

Features

  • Conditional execution based on runtime data evaluation
  • Sub-flow composition for hierarchical workflows
  • LLM resilience with automatic model switching
  • Complete retry strategy suite with configurable backoff policies

0.1.0 - 2025-01-24

Added

  • Initial release of Agent Orchestration Library
  • ExecutionContext for dependency injection
  • AgentBlock base class with validation and events
  • EventEmitter pub/sub system
  • Flow for multi-agent orchestration
  • ExponentialBackoffRetry strategy
  • Comprehensive documentation
  • Full test suite
  • Type hints throughout
  • MIT License

Core Features

  • Type-safe agent execution with Pydantic
  • Event-driven architecture
  • Dependency injection container
  • Sequential and parallel execution
  • Retry strategies
  • Progress tracking
  • Error handling