Lár v1.6.0: Memory Compression & Economic Constraints

Lár v1.6.0: Memory Compression & Economic Constraints

We are excited to announce Lár v1.6.0, introducing powerful new defensive mechanisms for mission-critical Agents.

The Problem

As Agents become more autonomous (like the DynamicNode introduced in v1.5), they run the risk of two things:

1. Context Bloat: Parallel workers returning huge documents that overwhelm the downstream context window (The "Black Hole").
2. Infinite Loops: An agent getting stuck in a self-correction loop and burning through API credits.

The Solution

1. Memory Compression (ReduceNode)

The new ReduceNode natively supports the Map-Reduce pattern to prevent context window bloat. It reads multiple heavy documents from the GraphState, generates a concise executive summary, and then explicitly deletes the raw data keys from memory. This ensures the state "baton" remains light and prevents downstream models from choking or hallucinating on massive context sizes.

2. Economic Constraints (Token Budgets)

You can now set a strict, mathematical ceiling on an agent's execution cost. Simply add a token_budget integer to the initial_state. Every time an LLMNode runs, it deducts its usage from this budget. If the budget hits zero, the node refuses to execute. You can now mathematically guarantee an agent will never exceed a specific dollar amount, regardless of how complex its dynamic graph becomes.

3. Node Fatigue (Infinite Loop Protection)

The GraphExecutor now tracks precisely how many times a node is visited during a single execution trace. If an agent loops over the same node (like a "Fix Code" node) more times than the max_node_fatigue limit (default: 20), the executor trips a circuit breaker and safely terminates the run, logging a fatigue error.

New Showcase Example

• Added examples/advanced/11_map_reduce_budget.py to demonstrate ReduceNode, map-reduce parallel fan-out, and token budget exhaustion.

Minor Fixes

• Added ReduceNode to lar/__init__.py for top-level import.