v0.7.4

[v0.7.4] - 2026-02-01

Added

Examples

• Adaptive Waiting Example - Added example demonstrating adaptive waiting patterns for agents
• Nativity Playbook - New nativity.pb example playbook

Agent System

• Unified Message Batching - Implemented unified RollingMessageCollector for batching all agent messages including direct and meeting broadcasts. Human messages now trigger immediate flush of pending batches for responsive interaction

Performance

• FanoutCache for Concurrent Access - Added FanoutCache to handle concurrent agent access patterns more efficiently

Improved

LLM Context Management

• Accurate Token Counting - Refactored token counting to use tiktoken library for precise token calculations. Replaced rough estimations with accurate counts via get_messages_token_count and get_token_count utility functions
• Trigger Evaluation Examples - Refined trigger evaluation and yield examples in prompts for better clarity on computable vs non-computable triggers
• Final Instructions - Revised InterpreterPrompt final instructions to emphasize concise thinking, strict adherence to available playbooks, and clear communication of assumptions for more reliable expert-like behavior

Changed

Configuration

• BGN Playbooks Visibility - BGN playbooks are no longer made public by default for better security and encapsulation
• Message Batching Config - Updated message batching configuration and package dependencies
• Optional Return Values - Return values are now optional in playbooks for more flexible execution patterns

Development

• Stop Tracking .llm_cache_test - Removed .llm_cache_test folder from git tracking

Documentation

• README Updates - Updated README.md with latest information

Fixed

Observability

• LangFuse Message Spans - Fixed LangFuse spans for messages not displaying message content. Now skips LangFuse spans for EOM (End of Message) markers
• LangFuse Token Counts - Fixed LangFuse LLM calls not showing token counts. Introduced _llm_input_tokens dictionary to correlate input tokens between LLM start and end events with proper cleanup
• Streaming in Conditional Blocks - Fixed Say() calls inside generated if and else blocks not being streamed. Disabled streaming of Say() calls after indented code is detected to maintain execution consistency

Core Functionality

• Playbooks Attribute Check - Fixed potential recursion issues in BaseAgent by replacing hasattr() with direct dict access when checking for 'playbooks' attribute

Dependencies

• Lazy Imports - Refactored imports in messaging_mixin.py and meeting_manager.py to use lazy imports, resolving circular dependency issues