erwinmsmith/BEAM
BEAM — a MAS efficiency toolkit that trims unnecessary agent communication to save tokens without sacrificing quality.
What's novel
BEAM — a MAS efficiency toolkit that trims unnecessary agent communication to save tokens without sacrificing quality.
Code Analysis
9 files read · 3 roundsA framework that wraps multi-agent LLM pipelines in a graph structure and learns which agent-to-agent edges to keep by training n² scalar logits via REINFORCE-style policy gradient, with three thin strategy variants (prune, dropout, Bayesian) that share the same core training loop.
Strengths
Clean graph/node/optimizer separation with a flexible LLM abstraction layer and registry pattern; the spatial/temporal edge distinction and topological execution model are well-structured and easy to extend with custom agent functions.
Weaknesses
No tests exist, the 'Bayesian' strategy never actually fits its Gaussian Process and the core training is standard REINFORCE rather than true Bayesian optimization, and the three strategies are largely redundant copies of the same policy-gradient loop with minor variations.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
16
Contributors
3
Files
41
Active weeks
6
Repository
Language
Python
Stars
3
Forks
1
License
—