Tag
This paper empirically analyzes the cost-effectiveness of code execution in LLM-based program repair agents, finding that execution is used heavily but often indiscriminately, and that restricting execution can save significant cost with minimal impact on repair success.
SelfCompact is a scaffolding approach that lets language models autonomously decide when and how to compact long agent traces, achieving better performance with reduced token costs compared to fixed-interval methods.
A quick `hf sync` command helps sync agent session traces so they are never lost and can be used across machines.
Harrison Chase announces a post-trained model for detecting issues in production agent traces, claiming SOTA accuracy at 10-100x cheaper rates than frontier models.
LangChain engineers detail how they built a custom inverted index for SmithDB to support full-text search and JSON filtering over large agent traces stored in object storage, achieving median latency of 400ms despite huge payload sizes.
Introduces RedAct, a framework to protect agent traces from procedural skill leakage by selectively redacting sensitive details while preserving audit evidence, along with CapTraceBench for evaluation.
TRL now supports fine-tuning models on agent traces from various sources like Claude Code, Codex, OpenClaw, and Pi, moving towards a standardized stack for training agentic models.
Introduces Reasoning Primitive Induction, a method that mines successful ReAct traces to cluster recurrent reasoning moves into typed pseudo-tools, outperforming the original agent by tens of percentage points on benchmarks.
New documentation page for rendering Agent Traces on the Hugging Face Hub has been released.
A library called teich converts agent traces into supervised fine-tuning (SFT) datasets, simplifying dataset preparation for AI training.
ZenithDB is a new open-source Rust database designed for storing and querying AI agent traces. It achieves sub-millisecond trace fetch latency at 1 billion rows by colocating all spans of a trace during compaction, and includes embedded full-text search and late materialization.