Reliable Code, Durable Sessions, and Self-Growing Skills

Research

Monte Carlo Tree Search Boosts Agent Code Generation

Introduces "Tree Search for LLM Agent Code Generation" (TAG), a training-free method using Monte Carlo Tree Search to optimize agent-written code for tool use. It suggests we can significantly improve agent reliability without retraining by searching over multiple code solutions.

Agent-as-Judge: Scalable Self-Evaluation

Introduces "Agent-as-Judge," where the agent evaluates its own performance via comparative feedback. This offers a scalable alternative to expensive human oversight for long-horizon or complex tasks.

Analysis

Autonomous Skill Discovery for Self-Growing Agents

Proposes "Agentic Skill Discovery" for agents to autonomously build and reuse skill libraries from past trajectories. This is the shift from one-off coding to building reusable, self-growing automation building blocks.

Multi-Agent Collaboration Under Uncertainty

Analyzes "adaptive exploration" in multi-agent systems, where agents dynamically hunt for information to resolve uncertainty. Useful for designing collaborative systems that don't just banish but actually work together.

Tools

Durable Agent State with LangGraph Sessions

LangGraph adds robust session management, solving the headache of persisting agent state across crashes and restarts. If you are building stateful agents in production, this is a critical primitive for reliability.