The Planning Test, Durable State, and Why Agent Teams Fail

New Benchmark for Long-Horizon Planning in AI Agents

Paper introduces a new benchmark for long-horizon agent planning, just dropped on arxiv. This is fundamental for measuring real-world capability, not just short-term task completion.

autonomous Tool Learning

Tackles the problem of agents learning to improve their own tool use over time. If you're building systems that need to get better without constant retraining, this is the direction.

Stateful Session Management in LangGraph

Directly addresses how to build agent memory that survives across sessions. Critical for building anything beyond a simple one-shot bot.

Analyzing Multi-Agent Failure Patterns

Analyzes failure modes in multi-agent systems. Understanding why teams of agents break is key to building more reliable orchestration layers.

Self-Improvement Through Agent Self-Evaluation

Proposes a framework for agents to self-evaluate their own output loops. Moves past simple reward signals toward more nuanced self-assessment for iterative tasks.

Dynamic Tool Use Selection for Agents

Looks at how to make agent tool use more dynamic and context-dependent, rather than relying on a fixed set. Tangential to tool-learning but focuses on the selection mechanism.