Repository navigation
0.4.0
·
428 commits
to master
since this release
CloudLLM 0.4.0: Introducing Mullti-Agent Council Support and basic (alpha) MCP Tool Support
Deploy compute to as many agents as you can afford and organize them under different topologies:
- Parallel Mode - Independent Expert Analysis
- Round-Robin Mode - Sequential Deliberation
- Moderated Mode - Expert Panel with Chair
- Hierarchical Mode - Multi-Layer Problem Solving
- Debate Mode - Adversarial Refinement with Convergence
We've also introduced early support for Tool use, we'll be focusing heavily on tool usage for agents on our next releases.
See the TUTORIAL and examples
0.4.0 OCT/13/2025
- feat: Multi-Agent Council System
- New council.rs module implementing collaborative multi-agent orchestration
- Five council modes for different collaboration patterns:
- Parallel: All agents process prompt simultaneously, responses aggregated
- RoundRobin: Agents take sequential turns building on previous responses
- Moderated: Agents submit proposals, moderator synthesizes final answer
- Hierarchical: Lead agent coordinates, specialists handle specific aspects
- Debate: Agents discuss and challenge each other until convergence
- Agent identity system with name, expertise, personality, and optional tool access
- Conversation history tracking with CouncilMessage and round metadata
- CouncilResponse includes final answer, message history, rounds executed, convergence score, and total tokens used
- feat: Tool Protocol Abstraction Layer (tool_protocol.rs)
- Flexible abstraction for connecting agents to various tool protocols
- ToolProtocol trait with execute(), list_tools(), get_tool_metadata()
- Support for MCP (Model Context Protocol), custom functions, and user-defined protocols
- ToolResult struct with success status, output, error, and execution metadata
- ToolParameter system with support for String, Number, Integer, Boolean, Array, Object types
- ToolMetadata with parameter definitions and protocol-specific metadata
- ToolRegistry for centralized tool management
- Tool and ToolError types for type-safe tool operations
- feat: Tool Adapters (tool_adapters.rs)
- CustomToolAdapter: Execute user-defined Rust closures as tools
- MCPToolAdapter: Integration with Model Context Protocol servers
- OpenAIToolAdapter: Compatible with OpenAI function calling format
- All adapters implement async ToolProtocol trait
- feat: Automatic Tool Execution in Agent Generation
- Agents automatically discover and execute tools during response generation
- Tool information injected into system prompts with name, description, and parameters
- JSON-based tool calling format: {"tool_call": {"name": "...", "parameters": {...}}}
- Automatic tool execution loop with max 5 iterations to prevent infinite loops
- Tool results fed back to LLM as user messages for continued generation
- Token usage tracked cumulatively across all LLM calls and tool executions
- New AgentResponse struct returns both content and token usage
- Agent::generate_with_tokens() method for internal use with token tracking
- feat: Convergence Detection for Debate Mode
- Jaccard similarity-based convergence detection for debate termination
- Compares word sets between consecutive debate rounds
- Configurable convergence threshold (default: 0.75 / 75% similarity)
- Early termination when agents reach consensus, saving tokens and cost
- Convergence score returned in CouncilResponse for inspection
- calculate_convergence_score() and jaccard_similarity() helper methods
- feat: Token Usage Tracking in Council Modes
- Parallel mode tracks and aggregates tokens from all concurrent agents
- RoundRobin mode accumulates tokens across sequential turns
- Token usage includes all LLM calls plus tool execution overhead
- CouncilResponse.total_tokens_used provides complete cost visibility
- feat: Comprehensive Multi-Agent Tutorial (COUNCIL_TUTORIAL.md)
- Cookbook-style tutorial with progressive complexity
- Five detailed recipes demonstrating each council mode
- Real-world carbon capture strategy problem domain
- Examples use multiple LLM providers (OpenAI, Claude, Gemini, Grok)
- Up to 5 agents per example with distinct expertise and personalities
- Tool integration example with MCPToolAdapter
- Best practices guide for agent design and council mode selection
- Troubleshooting section with common issues and solutions
- Complete multi-stage pipeline example combining multiple modes
- test: Added comprehensive test coverage
- test_agent_with_tool_execution: Validates tool discovery, execution, and result integration
- test_debate_mode_convergence: Validates convergence detection with mock agents
- test_parallel_execution: Tests concurrent agent execution
- test_round_robin_execution: Tests sequential turn-taking
- test_moderated_execution: Tests proposal aggregation
- test_hierarchical_execution: Tests lead-specialist coordination
- test_debate_execution: Tests debate discussion flow
- All 12 tests passing
- refactor: Code quality improvements
- Fixed all compiler warnings
- Switched from std::sync::Mutex to tokio::sync::Mutex for async compatibility
- Removed unused imports and assignments
- Improved error handling in council execution paths
- docs: Added inline documentation for council system and tool protocol
- Module-level documentation with architecture diagrams
- Example code in doc comments
- Detailed parameter and return value documentation