Lár v1.1 Release Notes: "The Metacognition Update"

Lár v1.1 Release Notes: "The Metacognition Update"

We are proud to announce Lár v1.1, the "Metacognition Update". This release introduces Level 4 Agency capabilities, allowing agents to modify their own execution logic at runtime while maintaining strict "Glass Box" compliance.

Major Features

1. Dynamic Graphs (Metacognition)

Agents can now inspect and modify their own topology during execution.

• DynamicNode: A new primitive that allows a node to return a new subgraph or topology modification instead of just a state update.
• TopologyValidator: A deterministic validation engine that ensures any self-assembled graph is valid, safe, and terminates ensuring compliance with safety constraints.
• Audit-Logged Self-Modification: Every change to the graph structure is recorded in the State-Diff Ledger, preserving the "Glass Box" guarantee even for self-modifying code.

2. New Examples & "Metacognition" Labs

We have added 5 advanced examples demonstrating dynamic graph capabilities:

• 25_dynamic_depth.py: An agent that decides how deep to think based on the question's complexity.
• 26_tool_inventor.py: An agent that writes its own Python tools on the fly to solve novel problems.
• 27_self_healing.py: A pipeline that detects failures (e.g., auth errors), dynamically inserts a "fix" node (e.g., credential rotation), and resumes execution.
• 28_adaptive_deep_dive.py: An agent that spawns parallel sub-agents to research topics it identifies as important.
• 29_expert_summoner.py: An agent that dynamically instantiates specialized "expert" nodes based on the user's query topic.

3. Compliance & Safety Updates

• EU AI Act & FDA Compatibility: Updated compliance documentation to address "Self-Modifying AI". By using TopologyValidator and the immutable State-Diff Ledger, Lár v1.1 ensures that even dynamic behavior is fully traceable and reproducible.
• Safety First: New documentation alerts and guidelines for deploying dynamic agents in production environments.

Improvements

• Example Reorganization: All examples have been moved to lar/examples/ and organized into clear categories:
  • basic/: Fundamental primitives (Linear, Branching).
  • patterns/: Common agentic patterns (RAG, ReAct, Map-Reduce).
  • compliance/: Safety and auditing examples (Human Jury, Audit Logs).
  • scale/: High-performance and distributed examples.
  • metacognition/: The new v1.1 dynamic graph examples.
• Website Updates:
  • Red Teaming Hook: Added a prominent hook to the Compliance section of the homepage.
  • Doc Improvements: Fixed documentation rendering for alerts/admonitions and audited all links.
  • Metacognition Section: Added a dedicated section for dynamic graphs.
• Version Bump: Framework version updated to 1.1.0.

Upgrading

To upgrade to Lár v1.1, simply run:

pip install lar-engine==1.1.0

Documentation

• New Metacognition Guide
• Compliance & Safety